Somewhere around the fourth article in this series, a pattern started to feel less like coincidence and more like a structural feature of how we practice.

Every one of these seven patients - the faller, the GI bleeder, the stented AF patient, the dialysis patient, the 140-kilogram patient, the one who stroked on a clean apixaban level, the one with a tumor quietly rewriting its own coagulation rules - got treated, at some point in the last two decades, by a rule that nobody could quite trace back to a trial. Not because clinicians were careless. Because the evidence simply wasn’t there yet, and a specialty that runs on protocols doesn’t tolerate a vacuum for long. Something fills it. Usually caution. Usually a round number - 120 kg, CrCl 30, “wait a week,” “switch the drug.” And once that number sits in a guideline table for a few years, it stops looking like a placeholder and starts looking like knowledge.

What this series has tried to do, one phenotype at a time, is go back and check which of those numbers were ever actually earned.

When absence gets mistaken for danger

Three of the seven articles turn on almost the identical mechanism, even though they live in completely different organ systems.

When Falling Doesn’t Mean You Should Stop the Anticoagulant opened the series with the clearest version of the pattern. An elderly patient falls, and anticoagulation quietly gets deprioritized - not because a trial showed falls predict fatal bleeding, but because the logic felt intuitive enough that nobody demanded the trial. BAFTA and the Donzé cohort did the arithmetic that the reflex had skipped: a patient on warfarin needs to fall roughly 295 times in a single year before the bleeding risk from those falls outweighs the stroke protection the drug provides. The fear was real. The math behind it was never there.

The 120-Kilogram Line We Drew From Silence, Not Data told a version of the same story with a number attached. Ask a room of cardiologists where the DOAC weight cutoff comes from and most will say “BMI over 40” with total confidence - and almost none of them can name the trial that established it, because no trial did. The 2016 ISTH guidance was honest about this: it wasn’t reporting a safety signal, it was reporting a gap, and it filled that gap with caution because that’s what you do with a gap. By 2021, apixaban and rivaroxaban subgroup data and a fresh guidance update had quietly closed most of it. The threshold didn’t get disproven so much as it got outgrown.

And When the Gut Bleeds and the Heart Still Needs Protecting ran the pattern in reverse. Here the instinct wasn’t premature restriction - it was premature caution about restarting. A patient bleeds, the endoscopist fixes the bleeder, and the reflex is “not yet, maybe never.” Witt’s cohort and Staerk’s much larger Danish registry both pointed the same direction: restarting anticoagulation after a GI bleed cut mortality by roughly 60 to 70 percent, with only a modest and often nonsignificant bump in rebleeding. The stroke that might happen “later, somewhere else” turned out to be far more dangerous than the bleed that had already been treated.

Three organ systems, three completely different patient phenotypes, one identical epistemic error: mistaking the absence of reassuring data for the presence of a real risk.

Subscribe now

When every intervention creates a new problem

The next cluster of articles is a different kind of hard - not fear built on silence, but genuine, structural risk-stacking, where every drug you add to protect against one thing raises your exposure to another.

When Every Drug Makes Something Worse: Anticoagulation in AF Patients After PCI is the cleanest example. Stroke risk demands anticoagulation. Stent thrombosis demands antiplatelet therapy. For years the default answer to that collision was triple therapy, warfarin, aspirin, and a P2Y12 inhibitor, stacked together because each piece had its own clear rationale and nobody had seriously tested whether the combination itself was the problem. WOEST asked that question in 2013 and got an answer nobody expected: dropping aspirin cut bleeding almost in half without any signal of increased ischemic events. AUGUSTUS, with its elegant 2×2 factorial design, then separated the two questions that had been tangled together the whole time, anticoagulant choice and antiplatelet choice, and found that apixaban plus a single P2Y12 inhibitor, no aspirin, was the safest regimen tested, with fewer hospitalizations and deaths than the VKA-based comparison. The triple-therapy reflex wasn’t wrong because someone miscalculated. It was wrong because nobody had isolated which component was actually buying protection and which was only adding hazard.

The Renal Cliff: Why Our Confidence in Anticoagulation Collapses Exactly Where Patients Need It Most describes a version of risk-stacking that evidence simply hasn’t caught up with yet. CKD stage 3 is comfortable territory - the pivotal DOAC trials all enrolled enough of these patients to speak with real confidence. Cross into dialysis, though, and the guideline confidence that felt so solid a few paragraphs earlier just stops. RENAL-AF, AXADIA-AFNET 8, and Valkyrie are each too small to be definitive on their own - underpowered, stopped early, or built around modest cohorts. Two of the three leaned toward DOACs on safety: AXADIA’s point estimate favored apixaban, though it never reached non-inferiority, and Valkyrie showed a significant halving of major bleeding with rivaroxaban. RENAL-AF’s own point estimate actually tilted the other way, toward more bleeding on apixaban, with a confidence interval too wide to mean much on its own. What made the three worth reading together wasn’t unanimous safety data - it was that not one of them, including the trial that leaned toward warfarin, produced a signal strong enough to justify defaulting to it. This is the one article in the series where the honest conclusion isn’t a corrected rule. It’s an acknowledgment that the rule genuinely doesn’t exist yet, and that clinicians are left triangulating from imperfect trials because nothing better has been built.

When the drug is working and the patient still isn’t protected

The final two articles are the hardest, because they don’t resolve into a corrected reflex or a closing evidence gap. They describe patients where anticoagulation is doing exactly what it’s supposed to do, and something still gets through.

The Patient Who Strokes Anyway starts from an uncomfortable statistic: something close to a quarter to a third of AF patients who present with ischemic stroke are already anticoagulated at the time. Seiffge’s pooled registry data made the scale of this visible for the first time; Polymeris’s etiological work showed that maybe a third of these events trace back to genuine underdosing or poor adherence - but close to another third have nothing to do with the anticoagulant at all, and the remainder occur in patients who are, by every measurable standard, adequately treated. The habitual response - switch the drug, add an antiplatelet, tighten counseling - turns out to rest on very little evidence. What the ELAN and OPTIMAS trials actually demonstrated was almost the opposite of caution: earlier restart after a breakthrough stroke, not later, is what the data supports. And for the genuinely unexplained cases, left atrial appendage occlusion is emerging as a rational option worth raising early rather than a last resort.

When the Tumor Writes Its Own Rules for Clotting closes the series with the phenotype where the underlying biology is actively working against the drug rather than passively exposing its limits. Cancer patients are simultaneously the most thrombosis-prone and the most bleeding-prone population anticoagulation ever treats, and both risks come from the same source - a tumor shedding tissue factor and procoagulant material directly into the bloodstream, sustained by an inflammatory state that shifts with every chemotherapy cycle. CLOT established LMWH as the standard for fifteen years by cutting recurrent VTE nearly in half compared to VKA. Then Hokusai-VTE Cancer, SELECT-D, and CARAVAGGIO tested whether an oral DOAC could replace the daily injections, and CARAVAGGIO in particular showed apixaban matching LMWH’s efficacy without SELECT-D’s excess GI bleeding - but only after the field learned, the hard way, that gastrointestinal cancers change that calculus meaningfully. Most recently, API-CAT asked whether the full-dose DOAC even needs to continue past six months once initial treatment is complete, and found that a reduced dose held up for extended treatment. This is a population where the anticoagulant itself keeps getting redesigned around the tumor’s behavior, not the other way around.

Subscribe now

The thread underneath all seven

Line these seven patients up next to each other and the differences start to matter less than the shape they share. In every case, the field’s first response to a hard clinical question was a rule built on the safest available assumption - avoid the drug, delay the restart, default to triple therapy, hold off on dialysis patients - and in nearly every case, that rule outlived the honest uncertainty that had produced it. The correction, when it came, rarely arrived from a single landmark trial. It came from years of subgroup analyses, imperfect small RCTs, registry data, and guideline committees willing to revisit a threshold they themselves had set.

The two articles that don’t fit this arc as neatly - dialysis, and breakthrough stroke - are worth sitting with precisely because they don’t resolve. They’re a reminder that not every clinical reflex is waiting for its BAFTA or its AUGUSTUS. Some of them are waiting for a trial that hasn’t been designed yet, or for a mechanism - factor XI inhibition, appendage occlusion in the truly refractory patient - that might change the entire framing rather than just adjusting the threshold.

If there’s a single practical takeaway that survives all seven scenarios, it’s probably this: the caution that feels most defensible at the bedside - stopping the drug, delaying the restart, avoiding the “unstudied” patient - is exactly the decision most likely to have outrun its evidence. The patients who actually break the rules we’ve built around them are rarely the ones the rules were designed to protect. They’re the ones the rules never quite got around to studying.

—Cardio Concepts

Series index

1. When Falling Doesn’t Mean You Should Stop the Anticoagulant - elderly patients at risk of falls

2. When the Gut Bleeds and the Heart Still Needs Protecting - restarting anticoagulation after GI bleeding

3. When Every Drug Makes Something Worse: Anticoagulation in AF Patients After PCI - triple therapy and its dismantling

4. The Renal Cliff: Why Our Confidence in Anticoagulation Collapses Exactly Where Patients Need It Most- CKD and dialysis

5. The 120-Kilogram Line We Drew From Silence, Not Data - obesity and DOAC dosing

6. The Patient Who Strokes Anyway - breakthrough stroke on anticoagulation

7. When the Tumor Writes Its Own Rules for Clotting - cancer-associated thrombosis

Subscribe now

Explore the Cardio Concepts Library

There’s a paradox that anyone who has ever prescribed anticoagulation to a cancer patient has run into, even without always naming it: this is the population that is simultaneously the most prone to clotting and the most prone to bleeding on anticoagulant therapy. It isn’t the usual “balance two risks” problem you see in elderly patients or those with CKD. Here, both risks rise together, from the same source - the tumor itself.

Cancer cells don’t sit passively waiting for chemotherapy. They shed tissue factor and procoagulant microparticles directly into the circulation, and they trigger neutrophils to release NETs - neutrophil extracellular traps, webs of DNA that normally catch bacteria but here become scaffolding for clot formation. At the same time, inflammatory cytokines like IL-6 and TNF-α sustain a chronic systemic inflammatory state that damages the endothelium broadly, including the gastrointestinal mucosa, already fragile from the tumor itself or from chemotherapy. This is why cancer-associated thrombosis (CAT) can be thought of as “unstable thrombosis”: it doesn’t follow the classic Virchow model seen in AF or ordinary VTE. It runs on a continuously active inflammatory substrate that shifts with every chemotherapy cycle and every stage of disease.

Over the past two decades, medicine has had to rewrite the rules of anticoagulation for this population not once, but twice.

Why warfarin failed exactly where it once ruled

Before 2003, VKA was the default choice for VTE in cancer patients, simply because it was the only long-term option available. But anyone who has ever chased an INR in a patient undergoing chemotherapy knows the feeling of running after a number that never sits still: vomiting reduces vitamin K absorption, chemotherapy agents interact broadly, the liver may be compromised by metastases or drug toxicity, and patients are often too unstable to maintain an acceptable time-in-therapeutic-range.

CLOT (Lee et al., NEJM 2003) was the first study to precisely quantify the cost of that instability. In 676 patients with active cancer and proximal DVT or PE, long-term dalteparin compared with VKA cut recurrent VTE from 17% to 9% - HR 0.48, nearly halving recurrence risk, with major bleeding rates numerically similar between groups (6% versus 4%). This wasn’t a marginal, debatable difference in efficacy. It was strong enough to make LMWH the standard of care for CAT for more than fifteen years, forming the foundation for every guideline from ASCO to ACCP to ISTH.

But daily subcutaneous LMWH injections, sustained over many months, in a patient already enduring chemotherapy, radiation, and countless procedures - that’s a real burden. The natural next question was whether an oral DOAC could take its place.

The DOAC era opens - and immediately exposes a blind spot

Hokusai-VTE Cancer (Raskob et al., NEJM 2018) was the first answer, and it was both a win and a warning. In 1,050 patients, after an initial LMWH lead-in, oral edoxaban achieved non-inferiority against dalteparin on the composite of recurrent VTE and major bleeding - 12.8% versus 13.5%. But breaking the two components apart told a more complicated story: recurrent VTE fell (7.9% versus 11.3%), while major bleeding rose significantly, 6.9% versus 4.0%, HR 1.77, particularly pronounced in patients with gastrointestinal cancers.

SELECT-D (Young et al., JCO 2018), though only a pilot study with 406 patients, drew this signal even more sharply. Rivaroxaban cut recurrent VTE dramatically, 4% versus 11%, HR 0.43, but clinically relevant non-major bleeding (CRNMB) rose almost fourfold, HR 3.76, particularly pronounced in esophageal and gastro-esophageal junction cancers. The signal was strong enough that the trial had to pause enrollment of esophageal cancer patients partway through.

This is where the CAT story takes a turn that’s far more interesting than “are DOACs better than LMWH.” The real question became: which DOAC, for which tumor? These two trials, despite SELECT-D’s small size, played an irreplaceable role - they were the first to force clinicians to see gastrointestinal cancer not as a vague risk factor but as a specific relative contraindication to certain molecules.

CARAVAGGIO: when one molecule separates itself from the pack

If Hokusai and SELECT-D raised the question, CARAVAGGIO (Agnelli et al., NEJM 2020) was the trial that answered it systematically. A key design difference: CARAVAGGIO deliberately included patients with gastrointestinal cancer, rather than avoiding this high-risk group as earlier trials had, excluding only brain tumors.

In 1,170 patients, apixaban achieved non-inferiority against dalteparin on recurrent VTE (5.6% versus 7.9%, HR 0.63), which was unsurprising given prior data. What mattered was major bleeding: 3.8% versus 4.0%, HR 0.82 - no increase, even in the gastrointestinal cancer subgroup. This was the first time a DOAC had proven itself both effective and safe in exactly the population where other molecules had shown a clear weakness.

The result wasn’t merely academic. It directly reshuffled priority rankings across practice guidelines - ASCO 2023, ITAC 2022, and ASH 2021 all moved apixaban to first-line status in CAT, a consensus that rarely forms so quickly and so cleanly based on a single trial, however strong.

The meta-analysis by Mulder et al. (Blood 2020), pooling four RCTs including Hokusai-VTE Cancer, SELECT-D, and CARAVAGGIO, found that DOACs as a class trended toward lower recurrent VTE than LMWH - RR 0.68, not statistically significant. It also found no clear signal that any one DOAC outperformed another. But it confirmed the more important point: safety profiles differ meaningfully between molecules, and stratifying by cancer type is the crux of the decision, not a footnote to mention in passing.

Subscribe now

From treatment to prevention: a story barely told before 2019

Most discussion around CAT centers on treating VTE that has already occurred. But there’s a different, quieter question with arguably greater potential to reshape practice: can VTE be prevented altogether in high-risk ambulatory cancer patients?

CASSINI (Khorana et al., NEJM 2019) tested rivaroxaban 10 mg daily for 180 days in patients with a Khorana score ≥2. The primary endpoint across the full 180 days missed statistical significance - 6.0% versus 8.8%, p=0.10. But restricting the analysis to the period patients were actually taking the drug revealed a clear effect: 2.6% versus 6.4%, HR 0.40, p=0.007. This is one of those cases where a trial “negative” on its primary endpoint still carries real clinical weight - the gap between a drug’s intrinsic efficacy and its measured effectiveness in a population with high discontinuation rates, a phenomenon familiar to anyone who has worked with ambulatory cancer patients.

AVERT (Carrier et al., NEJM 2019), published the same year, reinforced the story with low-dose apixaban (2.5 mg twice daily), again in Khorana ≥2 patients. VTE fell sharply: 4.2% versus 10.2%, HR 0.41. But this time the cost was more visible: major bleeding rose significantly, 3.5% versus 1.8%, HR 2.00, p=0.046. NNT to prevent one VTE was 17; NNH to cause one major bleed was 59 - a benefit-risk ratio still tilted toward benefit, but close enough to demand careful patient selection rather than blanket application.

CASSINI and AVERT, read together, aren’t duplicate trials - they’re two halves of the same story: primary prophylaxis with low-dose DOACs is feasible and genuinely beneficial in Khorana ≥2 patients, but it requires an explicit benefit-risk calculation that the clinician has to make patient by patient, with no universal formula.

API-CAT: the question no one asked for two decades

There has been an unspoken assumption since the CLOT era: once you decide to extend CAT treatment beyond six months, full-dose anticoagulation is mandatory, because recurrence risk in active cancer is persistent and doesn’t decline over time the way it does in non-cancer VTE. No one had actually tested this assumption in an RCT until recently.

API-CAT (Mahé et al., NEJM 2025) did exactly that. In 1,766 patients who had completed at least six months of initial treatment, two-thirds with distant metastatic disease, the trial compared continuing reduced-dose apixaban (2.5 mg twice daily) against full dose (5 mg twice daily) for an additional 12 months. Result: recurrent VTE 2.1% versus 2.8%, meeting non-inferiority. But the more interesting finding was clinically relevant bleeding - meaningfully lower with the reduced dose, 12.1% versus 15.6%, subHR 0.75.

This is the first RCT evidence supporting dose individualization during the extended treatment phase of CAT - a concept that barely existed in clinical thinking before 2025, when the default was “full dose until treatment stops.” It raises a concrete practical question: for a metastatic cancer patient who has been stable for six months of VTE treatment, does maintaining full-dose anticoagulation indefinitely still make sense, or are we accepting unnecessary bleeding risk out of habit?

Subscribe now

What the evidence still hasn’t reached

The picture of CAT, though reasonably complete for most solid-tumor patients, still has significant blind spots that none of the studies in this overview resolve. Primary or metastatic brain tumors were systematically excluded from the major RCTs, CARAVAGGIO excluded them entirely, and the others followed suit, so data on DOAC safety in this population is essentially absent. Thrombocytopenia below 50×10⁹/L, common after myelosuppressive chemotherapy, also falls outside the scope of the pivotal trials. Drug interactions between DOACs and targeted therapies metabolized through CYP3A4/P-glycoprotein, a drug class occupying an increasingly large share of modern cancer regimens, still lack systematic interaction data. And hematologic malignancies, with a fundamentally different coagulation pathophysiology from solid tumors, have barely been addressed as a distinct category.

These aren’t minor footnotes to list under “limitations” for completeness. They reflect a real state of affairs: even after two decades of research, cancer-associated thrombosis is best understood in exactly the patients easiest to enroll in an RCT - and least understood in the most complex patients, the ones clinicians actually have to make decisions about every day without RCT-level evidence to rely on.

A path that might finally uncouple two risks that have always traveled together

The entire history of CAT treatment, from CLOT to API-CAT, can be summarized as a continuous search: how to reduce thrombosis without paying for it in bleeding, in a population where both risks arise from the same tightly linked inflammation-coagulation mechanism. Every molecule from VKA to LMWH to today’s DOACs acts on the shared coagulation pathway - the pathway both necessary for physiologic hemostasis and where pathologic thrombosis forms. That’s why nearly every gain in efficacy has come with some cost to safety, large or small.

Factor XI and XIa inhibitors, asundexian, milvexian, abelacimab, are being studied on a different working hypothesis: that the contact pathway, in which factor XI participates, plays a more important role in pathologic thrombosis than in normal physiologic hemostasis. If that hypothesis holds, it would be the first time medicine has a tool capable of separating two risks that have always run in parallel in cancer patients.

In a population where thrombosis is unstable because it’s fed by the tumor itself, and bleeding is unstable because it’s fed by that same inflammatory substrate, the question worth sitting with isn’t which DOAC is best for cancer today - it’s whether we’re approaching a new generation of anticoagulants that no longer forces a trade-off between the two, or whether that’s still just a mechanistic hope that hasn’t yet been proven at scale.

If you'd like a one-page visual summary of the key trials, decision pathway, and practical takeaways from this article, you can download the companion framework below.

SAME SOURCE, SPLIT FATES

—Cardio Concepts

Previous The Patient Who Strokes Anyway

Next The Hard Patients: What Seven High-Risk Anticoagulation Scenarios Have in Common

References

1. Lee AYY, Levine MN, Baker RI, Bowden C, Kakkar AK, Prins M, et al.; CLOT Investigators. Low-molecular-weight heparin versus a coumarin for the prevention of recurrent venous thromboembolism in patients with cancer. N Engl J Med. 2003;349(2):146–153. doi:10.1056/NEJMoa025313

2. Raskob GE, van Es N, Verhamme P, Carrier M, Di Nisio M, Garcia D, et al.; Hokusai VTE Cancer Investigators. Edoxaban for the treatment of cancer-associated venous thromboembolism. N Engl J Med. 2018;378(7):615–624. doi:10.1056/NEJMoa1711948

3. Young AM, Marshall A, Thirlwall J, Chapman O, Lokare A, Hill C, et al. Comparison of an oral factor Xa inhibitor with low molecular weight heparin in patients with cancer with venous thromboembolism: results of a randomized trial (SELECT-D). J Clin Oncol. 2018;36(20):2017–2023. doi:10.1200/JCO.2018.78.8034

4. Agnelli G, Becattini C, Meyer G, Muñoz A, Huisman MV, Connors JM, et al.; Caravaggio Investigators. Apixaban for the treatment of venous thromboembolism associated with cancer. N Engl J Med. 2020;382(17):1599–1607. doi:10.1056/NEJMoa1915103

5. Khorana AA, Soff GA, Kakkar AK, Vadhan-Raj S, Riess H, Wun T, et al.; CASSINI Investigators. Rivaroxaban for thromboprophylaxis in high-risk ambulatory patients with cancer. N Engl J Med. 2019;380(8):720–728. doi:10.1056/NEJMoa1814630

6. Carrier M, Abou-Nassar K, Mallick R, Tagalakis V, Shivakumar S, Schattner A, et al.; AVERT Investigators. Apixaban to prevent venous thromboembolism in patients with cancer. N Engl J Med. 2019;380(8):711–719. doi:10.1056/NEJMoa1814468

7. Mahé I, Carrier M, Mayeur D, Chidiac J, Vicaut E, Falvo N, et al.; API-CAT Investigators. Extended reduced-dose apixaban for cancer-associated venous thromboembolism. N Engl J Med. 2025;392(14):1363–1373. doi:10.1056/NEJMoa2416112

8. Mulder FI, Bosch FTM, Young AM, Marshall A, McBane RD, Zemla TJ, et al. Direct oral anticoagulants for cancer-associated venous thromboembolism: a systematic review and meta-analysis. Blood. 2020;136(12):1433–1441. doi:10.1182/blood.2020005819

   Share

   Leave a comment

   Explore the Cardio Concepts Library

A patient with atrial fibrillation, CHA₂DS₂-VASc of 4, apixaban 5 mg twice daily, no missed doses, renal function normal, no interacting drugs on the chart. And yet: a new ischemic stroke. The chart looks clean. The adherence looks perfect. The instinct in the room is almost always to look for someone to blame - the patient forgot a dose, the pharmacy filled the wrong strength, someone miscalculated the renal cutoff. Sometimes that instinct is right. Often it isn’t.

This scenario is not rare. It happens often enough that it has its own name, breakthrough stroke, and its own emerging literature, distinct from the general AF-stroke evidence base most of us built our mental models on. The uncomfortable part is not that anticoagulation occasionally fails. It’s that when we go looking for a single fixable cause, we frequently don’t find one. And the response most of us learned by habit, switch the drug, add an antiplatelet, tighten adherence counseling, turns out to have very little supporting evidence behind it.

A quarter of AF strokes happen on therapy

The scale of this problem only became clear once someone pooled the registries. Seiffge and colleagues combined individual patient data from seven prospective multinational cohorts, RAF, RAF-DOAC, CROMIS-2, SAMURAI, NOACISP, Erlangen, Verona, covering more than 4,000 patients admitted for acute ischemic stroke with known AF. A quarter to nearly a third of AF patients who present with ischemic stroke are already on oral anticoagulation at the time. These are not low-risk patients who got unlucky. They tend to carry higher CHA₂DS₂-VASc scores and more severe strokes than patients who were never anticoagulated, and their three-month risk of a second ischemic event runs around 7 percent - roughly double the recurrence rate seen after a first stroke in someone not on oral anticoagulation.

That last figure is the one worth sitting with. Breakthrough stroke isn’t a statistical footnote sitting quietly inside the larger AF-stroke population. It behaves like its own clinical entity, with its own trajectory and its own excess risk, and until fairly recently it had almost no dedicated evidence to guide what happens next.

Why “adequate” anticoagulation still fails

The obvious first question is mechanistic: what is actually going wrong. Polymeris and colleagues, working from the NOACISP-LONGTERM cohort of roughly 3,000 patients, tried to classify the underlying etiology rather than simply describe the event rate. Around a third of breakthrough strokes traced back to poor adherence or genuine underdosing, while close to another third had a competing cause that had nothing to do with the atrial fibrillation itself - large-vessel atherosclerosis, small-vessel disease, occult malignancy, patent foramen ovale. The remainder sat in a murkier middle ground: therapeutic anticoagulation, no obvious competing etiology, and a stroke that occurred anyway - plausibly reflecting resistant left atrial appendage thrombus or an atrial cardiomyopathy that has progressed independently of rhythm control or anticoagulation intensity.

What the same analysis found next is arguably more important than the taxonomy itself. Switching to a different DOAC or adding an antiplatelet agent did not reduce recurrence, and one-year recurrence risk remained in the range of 7 to 10 percent regardless. This is the finding that should have quietly retired a whole category of reflexive clinical decisions. There is no mechanistic reason apixaban should outperform rivaroxaban in a patient whose stroke happened because of poor adherence, and there is no reason adding aspirin should help a patient whose real problem is atrial cardiomyopathy. Etiology is heterogeneous. A single reflexive maneuver applied to all of it was never going to work, and the data confirm that it doesn’t.

So if switching drugs and stacking antithrombotics isn’t the answer, what actually moves the needle? Two questions turned out to matter more: when do you restart anticoagulation after the acute event, and is there a structural option that sidesteps the pharmacology altogether.

Subscribe now

The 1-3-6-12 rule was never really evidence

For over fifteen years, the timing of anticoagulation resumption after an acute ischemic stroke in AF was governed by a rule of thumb rather than a trial: wait 1 day after a TIA, 3 days after a mild stroke, 6 after moderate, 12 after severe. It was a sensible-sounding compromise, born out of a legitimate fear of hemorrhagic transformation, but it was never tested against an early-initiation strategy in a randomized design. Clinicians followed it because there was nothing better, not because anyone had shown it was the optimal window.

ELAN changed that. Fischer and colleagues randomized 2,013 patients across 103 centers in 15 countries to early initiation, within 48 hours for mild-to-moderate strokes, day 6–7 for severe ones, versus a later, guideline-consistent window. At 30 days, the composite outcome occurred in 2.9 percent of the early group versus 4.1 percent of the later group, a difference that did not reach statistical significance but leaned firmly in favor of early treatment, and recurrent ischemic stroke was roughly half as frequent in the early arm. Symptomatic intracranial hemorrhage was vanishingly rare in both groups, only two cases each. The safety signal that had justified fifteen years of caution simply didn’t materialize.

OPTIMAS then supplied the scale ELAN couldn’t. Werring and colleagues randomized 3,648 patients across roughly 100 UK centers, deliberately including patients with severe strokes and hemorrhagic transformation, the exact population the old rule was most protective of. Early initiation, within four days, met non-inferiority against later initiation at 7 to 14 days, with a 90-day primary outcome of 3.3 percent versus 3.4 percent and symptomatic intracranial hemorrhage rates of 0.6 percent versus 0.7 percent. Even in the patients the rule was designed to protect, early anticoagulation held up.

Two large, independent, well-conducted trials converging on the same answer is not a subtle signal. The 1-3-6-12 rule wasn’t wrong because someone made an error; it was simply never built on outcome data in the first place. It was a reasonable placeholder that outlived its evidentiary justification by over a decade, and ELAN and OPTIMAS together are what finally closed that gap.

When the pharmacology has already failed once, is there a mechanical answer?

Timing solves one problem, how quickly to restart therapy, but it does nothing for the patient whose stroke happened despite genuinely adequate anticoagulation, the roughly one-third to one-half of breakthrough strokes without a clean adherence or competing-cause explanation. For that group, the appendage itself becomes a reasonable target, independent of whichever anticoagulant is chosen next.

Maarse and colleagues, through the international STR-OAC LAAO collaboration, built a propensity-matched cohort of 433 patients who underwent percutaneous left atrial appendage occlusion after a breakthrough stroke, compared against a matched group who continued oral anticoagulation alone. Stroke incidence was 2.8 per 100 patient-years in the occlusion group versus 8.9 per 100 patient-years in the anticoagulation-only group, without an increase in major bleeding. A roughly two-thirds relative reduction, in exactly the population where pharmacological strategies had already been tried and had already failed.

This is observational data, not a randomized trial, and the usual caveats about selection into a procedural cohort apply. But it is the strongest evidence currently available for a genuinely mechanistic answer to a problem that adherence counseling and drug-switching cannot solve, and it’s substantial enough that dedicated randomized trials, CLOSURE-AF and STR-OAC, are now underway to test it properly. For clinicians managing a patient who strokes on well-documented therapeutic anticoagulation, LAAO is no longer a theoretical option sitting at the margins of the discussion. It has a number attached to it now, even if that number still needs a randomized trial to confirm.

Separating antithrombotic effect from bleeding risk

The furthest-out development is also the most conceptually interesting, because it attacks the problem from a completely different angle: what if the antithrombotic effect and the bleeding risk didn’t have to travel together at all.

AXIOMATIC-SSP tested milvexian, a factor XIa inhibitor, added on top of dual antiplatelet therapy after stroke or TIA - a setting where bleeding risk from adding anything is usually the first concern. Sharma and colleagues randomized 2,366 patients across a dose range of milvexian versus placebo for 90 days. The trial did not meet its primary endpoint of symptomatic stroke plus silent infarction on MRI, but doses of 25 to 100 mg twice daily showed roughly a 30 percent relative reduction in symptomatic stroke, without an increase in major bleeding or symptomatic intracranial hemorrhage.

A phase 2 trial that misses its composite primary endpoint is not, on its own, a reason to change practice - and it shouldn’t be read as one here. What makes it worth including in this discussion is narrower and more specific: the safety signal. Adding a potent antithrombotic on top of existing dual antiplatelet therapy, in a stroke population, without increasing hemorrhage, is not something the field has seen before from a drug targeting hemostasis. It’s the first real clinical hint that factor XI/XIa inhibition might genuinely decouple thrombosis prevention from bleeding risk, the two variables that have been locked together in every anticoagulant we’ve used to date. Phase 3 programs, LIBREXIA-STROKE for milvexian, OCEANIC-STROKE for asundexian, will determine whether that early signal survives at scale.

Subscribe now

What this actually changes

Put together, these five studies don’t converge on one new algorithm. They converge on something more useful: a map of where reflexive habits have outlived their evidence, and where genuinely new options are emerging to fill the gap those habits used to occupy.

The instinct to switch anticoagulants after a breakthrough event turns out to rest on very little. The instinct to wait cautiously before restarting therapy after the acute event turns out to rest on even less - a rule of thumb, not a trial, now directly contradicted by two of them. What does have real evidence behind it is less comfortable, because it demands more individualized thinking rather than a protocol: identify the mechanism where you can, restart early rather than late in almost every case, and consider appendage occlusion seriously - not as a last resort, but as a rational next step - in the patient whose stroke genuinely can’t be explained by a fixable pharmacological problem.

Somewhere in the next several years, factor XI inhibition may add a sixth option that doesn’t force a tradeoff between the two things we’ve never been able to separate. Until then, the honest answer for the patient sitting in front of you, therapeutic INR or clean DOAC levels and all, is that “adequately anticoagulated” was never quite the same claim as “protected.” The evidence is only now starting to tell us what to do with that distinction.

Free download: A one-page clinical framework for one of the most challenging situations in stroke prevention, when an ischemic stroke occurs despite therapeutic anticoagulation. Evidence-based, practical, and designed to help clinicians move beyond simply switching DOACs

Adequate ≠ Protected

—Cardio Concepts

Previous The 120-Kilogram Line We Drew From Silence, Not Data

Next When the Tumor Writes Its Own Rules for Clotting

References

1. Seiffge DJ, De Marchis GM, Koga M, Paciaroni M, Wilson D, Cappellari M, et al. Ischemic stroke despite oral anticoagulant therapy in patients with atrial fibrillation. Annals of Neurology. 2020;87(5):677–687. doi:10.1002/ana.25700

2. Polymeris AA, Schäfer J, Seiffge DJ, Lyrer P, Gebhardt A, Bonati LH, et al. Aetiology, secondary prevention strategies and outcomes of ischaemic stroke despite oral anticoagulant therapy in patients with atrial fibrillation. Journal of Neurology, Neurosurgery and Psychiatry. 2022;93(3):240–248. doi:10.1136/jnnp-2021-328391

3. Fischer U, Koga M, Strbian D, Branca M, Galm C, Paciaroni M, et al.; ELAN Investigators. Early versus later anticoagulation for stroke with atrial fibrillation. New England Journal of Medicine. 2023;388(26):2411–2421. doi:10.1056/NEJMoa2303048

4. Werring DJ, Dehbi HM, Ahmed N, Arram L, Best JG, Balogun M, et al.; OPTIMAS Investigators. Optimal timing of anticoagulation after acute ischaemic stroke with atrial fibrillation (OPTIMAS): a multicentre, blinded-endpoint, phase 4, randomised controlled trial. Lancet. 2024;404:1731–1741. doi:10.1016/S0140-6736(24)02197-4

5. Maarse M, Seiffge DJ, Werring DJ, Boersma LVA, Vos JA, Salinas P, et al. Left atrial appendage occlusion vs standard of care after ischemic stroke despite anticoagulation. JAMA Neurology. 2024;81(11):1150–1158. doi:10.1001/jamaneurol.2024.2882

6. Sharma M, Molina CA, Toyoda K, Bereczki D, Amarenco P, Ntaios G, et al. Safety and efficacy of factor XIa inhibition with milvexian for secondary stroke prevention (AXIOMATIC-SSP): a phase 2, international, randomised, double-blind, placebo-controlled, dose-finding trial. Lancet Neurology. 2024;23(2):150–161. doi:10.1016/S1474-4422(23)00403-9

   Share

   Leave a comment

Ask a room of cardiologists and pharmacists where the DOAC weight cutoff comes from, and most will answer with confidence: BMI over 40, or weight over 120 kg, switch to warfarin. Few will be able to tell you which trial established that number. That’s because no trial did. The cutoff was never a finding. It was an absence dressed up as a rule.

This is worth sitting with for a moment, because it explains almost everything that happened afterward in this corner of anticoagulation. When the International Society on Thrombosis and Haemostasis published its first guidance on DOAC use in obesity in 2016, the authors weren’t reporting a safety signal. They were reporting a gap. Phase 3 trials of dabigatran, rivaroxaban, apixaban and edoxaban had enrolled comparatively few patients above 120 kg, so the confidence intervals around that population were wide enough to drive a truck through. Faced with that uncertainty, the reasonable clinical instinct was caution: recommend against DOACs above the threshold, or at minimum, check peak and trough drug levels before trusting the dose.

That instinct was defensible in 2016. It became something else by 2021.

A guideline built on what wasn’t known

The ISTH SSC’s original position deserves more credit than it usually gets, because it was honest about its own limitations. It didn’t claim DOACs were unsafe in obesity - it claimed the evidence to say otherwise didn’t yet exist. The document leaned on pharmacokinetic reasoning and whatever fragments of phase 3 and retrospective data were available, and it landed on a threshold - BMI >40 or weight >120 kg - below which every DOAC was considered acceptable, and above which the recommendation shifted toward avoidance or, if a DOAC was unavoidable, drug-level monitoring.

For a clinician standing at the bedside of an 82 kg patient with AF, this was a fairly typical evidence gap. For one standing across from a 135 kg patient with the same rhythm and the same CHA₂DS₂-VASc score, it meant an entirely different medication, different monitoring burden, and, practically speaking, a return to INR checks and dietary vitamin K counseling that the DOAC era was supposed to have retired.

The uncomfortable part is that this threshold wasn’t derived from watching obese patients bleed or clot at higher rates on DOACs. It was derived from not having watched them at all, at least not in numbers large enough to trust. The guidance was, in the most literal sense, precautionary - and precaution, once written into a document with the authority of an international society, tends to calcify into practice long after the reasons for it have shifted.

The subgroup that quietly answered the question

Three years later, Hohnloser and colleagues went back into the ARISTOTLE dataset, 18,201 patients, apixaban versus warfarin, and pulled out the extremes of body weight: those under 60 kg and those over 120 kg, 982 patients in the latter group. This wasn’t a new trial designed to settle the obesity question. It was a post-hoc look at data that had already been collected, which is exactly why it matters: these were real trial participants, randomized and followed prospectively, who happened to sit outside the range the ISTH guidance had flagged as uncertain.

What they found didn’t support the caution embedded in the 2016 guidance. Apixaban’s efficacy against stroke and systemic embolism, and its effect on mortality, held consistently at both extremes of body weight, the lightest patients and the heaviest, with no meaningful drop-off above 120 kg. The bleeding signal, if anything, ran in the opposite direction from what the obesity caution would have predicted - the clearest safety advantage for apixaban over warfarin appeared at the low end of body weight, not the high end, though the benefit remained consistent above 120 kg as well, with a statistically significant interaction across the weight spectrum.

This is the kind of finding that should reshape a threshold, because it directly tests the assumption the threshold was built on. If extreme body weight altered apixaban’s pharmacokinetics enough to compromise efficacy or safety, a large randomized dataset stratified by weight is exactly where that would surface. It didn’t.

Subscribe now

What the real world added

Randomized trial subgroups are convincing, but they’re still trial patients - generally healthier, more closely monitored, and less complex than the people pharmacists see filling prescriptions on a Tuesday afternoon. Coons and colleagues addressed that gap with a retrospective, matched cohort from UPMC: 1,840 patients with acute VTE, comparing DOACs against warfarin in patients weighing more than 100 kg, up to nearly 300 kg.

Recurrent VTE at 12 months didn’t differ between the groups. Major bleeding did - but not in the direction the 2016 threshold anticipated. Bleeding occurred in 0.5% of DOAC patients versus 2.4% of warfarin patients, a difference that didn’t reach statistical significance given the sample size, but pointed unmistakably toward DOACs being the safer option, not the riskier one, even in a population heavier and more heterogeneous than any phase 3 trial had captured.

By 2020, then, the picture had two independent lines of evidence, one randomized, one real-world, both suggesting that the original obesity threshold had been drawn more conservatively than the biology warranted.

The guidance catches up

The ISTH SSC’s 2021 update is where the story turns. Rather than defending the original cutoff, the committee revised it in light of accumulated evidence from EINSTEIN, AMPLIFY, Hokusai-VTE, ARISTOTLE, and the real-world cohorts published since 2016. The new recommendation is more permissive, but also more specific than the blanket caution it replaced: standard-dose apixaban or rivaroxaban is now recommended even above BMI 40 or 120 kg, while dabigatran, edoxaban, and betrixaban remain outside the recommendation at that extreme. Routine drug-level monitoring, a fixture of the 2016 guidance, was dropped as a requirement.

That drug-by-drug distinction matters clinically, and it’s easy to miss if you remember only the headline that “DOACs are now fine in obesity.” They aren’t uniformly fine - apixaban and rivaroxaban earned their place through the specific datasets described above, while dabigatran and edoxaban simply haven’t accumulated the same weight of evidence at the extremes. Prescribing habits built on the 2016 blanket caution, or built on the mistaken assumption that the 2021 update applies equally across the DOAC class, both miss what actually changed.

There’s a broader pattern here worth naming: evidence didn’t get stronger for one drug because that drug is pharmacologically superior in obesity. It’s tempting to read this as apixaban getting stronger because it happened to be the one large trial that ran the subgroup analysis, and because retrospective cohorts happened to capture DOACs as a class without always breaking down by molecule. The 2021 guidance reflects where the data landed, not necessarily where the underlying pharmacology sits - a distinction that matters when a patient in front of you doesn’t fit neatly into the population that generated the recommendation.

Where the evidence still runs out

Two gaps remain, and neither has closed since 2021.

The first is the upper tail of body weight. Almost everything described above, the ARISTOTLE subanalysis, the UPMC cohort, the guidance itself, clusters its “extreme” definition around 120 kg or BMI 40. Patients with BMI above 50 are represented sparsely enough that no randomized dataset can speak to them with confidence. The guidance extends standard dosing upward in principle, but the further a patient sits from the population actually studied, the more that recommendation is extrapolation rather than evidence.

The second is the post-bariatric surgery patient - a population that’s grown substantially as metabolic surgery has become more common, and one whose relevant physiology (altered absorption, rapidly changing weight) is different in kind from stable obesity rather than just a different point on the same curve. The 2021 guidance advises avoiding DOACs for at least four weeks after bariatric surgery, a pragmatic precaution rather than a data-driven threshold, because absorption in the immediate post-surgical period is simply not well characterized for any of these drugs.

Subscribe now

The gap that hasn’t closed

By 2023, ACC/AHA/ACCP/HRS guidance had folded this evidence into a formal recommendation, giving DOACs a Class IIa - reasonable to choose DOACs over warfarin in AF patients with severe obesity. On paper, the uncertainty that justified caution in 2016 has been substantially resolved for two of the four available drugs, in the population most commonly seen in clinic.

What hasn’t resolved at the same pace is prescribing behavior. Practice surveys continue to show clinicians more hesitant than the current guidance would suggest - a hesitancy that made complete sense in 2016 and makes progressively less sense with each passing year, yet persists anyway. Guidelines update in a single publication. Clinical habits update one prescriber, one patient encounter, at a time, and the two clocks rarely run at the same speed.

The 120-kilogram line was never really a biological boundary. It was a marker of what a handful of trials happened to enrol at the time they were designed. Watching how quickly it moved once someone bothered to look - and how slowly practice has followed - says less about obesity and anticoagulation specifically than it does about how much of what we treat as clinical certainty is really just the shape of the dataset we happened to have.

To accompany this article, I've made the complete one-page Cardio Concepts framework available as a free download, showing how the evidence behind the 120-kg threshold evolved, and where its boundaries still remain.

The 120-Kilogram Line

—Cardio Concepts

Previous The Renal Cliff: Why Our Confidence in Anticoagulation Collapses Exactly Where Patients Need It Most

Next The Patient Who Strokes Anyway

References

1. Martin K, Beyer-Westendorf J, Davidson BL, Huisman MV, Sandset PM, Moll S. Use of the direct oral anticoagulants in obese patients: guidance from the SSC of the ISTH. J Thromb Haemost. 2016;14(6):1308–1313. doi:10.1111/jth.13323

2. Hohnloser SH, Fudim M, Alexander JH, Wojdyla DM, Ezekowitz JA, Hasenfuss G, et al. Efficacy and safety of apixaban versus warfarin in patients with atrial fibrillation and extremes in body weight. Circulation. 2019;139(20):2292–2300. doi:10.1161/CIRCULATIONAHA.118.037955

3. Coons JC, Albert L, Bejjani A, Iasella CJ. Effectiveness and safety of direct oral anticoagulants versus warfarin in obese patients with acute venous thromboembolism. Pharmacotherapy. 2020;40(3):204–210. doi:10.1002/phar.2369

4. Martin KA, Beyer-Westendorf J, Davidson BL, Huisman MV, Sandset PM, Moll S. Use of direct oral anticoagulants in patients with obesity for treatment and prevention of venous thromboembolism: updated communication from the ISTH SSC Subcommittee on Control of Anticoagulation. J Thromb Haemost. 2021;19(8):1874–1882. doi:10.1111/jth.15358

5. Joglar JA, Chung MK, Armbruster AL, et al. 2023 ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. *Circulation*. 2023;149(1):e1–e156. doi:10.1161/CIR.0000000000001193

   Share

   Leave a comment

   Explore the Cardio Concepts Library

There’s a particular kind of discomfort that shows up on dialysis wards. A patient with atrial fibrillation, a CHA₂DS₂-VASc score comfortably above the threshold for treatment, and a chart that offers almost nothing to guide the decision that follows. We know how to dose apixaban in a patient with CrCl of 45. We know what to do at 35. Somewhere between stage 4 CKD and the dialysis chair, though, the evidence base that felt so solid a few paragraphs earlier in the guideline simply stops.

This isn’t a minor gap. Dialysis patients carry some of the highest stroke risk and some of the highest bleeding risk of any population we anticoagulate, and for years the honest answer to “what does the evidence say” was: not much. What’s changed recently isn’t that the uncertainty disappeared. It’s that we finally have enough small, flawed, genuinely informative trials pointing in roughly the same direction that ignoring them starts to feel like its own kind of risk.

The Territory We Already Know

Start with what isn’t controversial. In CKD stage 3, roughly CrCl 30–50 mL/min, the pre-specified subgroup analyses from the four pivotal DOAC trials tell a remarkably consistent story: direct oral anticoagulants hold their own against warfarin, and in several analyses do measurably better, particularly on intracranial hemorrhage. This is the part of the renal spectrum where most cardiologists already feel comfortable, where the guidelines speak with confidence, and where DOAC prescribing has become close to default practice.

The interesting part of the story begins one stage further down, at the point where that confidence should, in theory, start to erode.

Stage 4: The Turning Point

CKD stage 4, CrCl 25 to 30 mL/min, sits in an odd position. It’s not dialysis, so it doesn’t trigger the reflexive caution that end-stage renal disease does. But it’s low enough that many clinicians quietly default to warfarin anyway, on the theory that if we don’t have strong DOAC data here, the safer choice is the drug we’ve used for seventy years.

Stanifer and colleagues tested that instinct directly, using a pre-specified subgroup of ARISTOTLE limited to patients with CrCl 25–30. It’s a small slice of a large trial, 269 patients, 136 on apixaban, but the signal inside it is not subtle. Major bleeding was reduced by roughly two-thirds (HR 0.34, 95% CI 0.14–0.80), and the composite of major plus clinically relevant non-major bleeding showed a similar magnitude of benefit (HR 0.35, 0.17–0.72). There was even a suggestion, the authors are appropriately careful about how far to push this given the sample size, that the bleeding advantage of apixaban actually grows as renal function worsens from >30 to the 25–30 range, rather than shrinking.

That’s counterintuitive if your mental model is “less kidney function, more caution needed with DOACs.” It fits better with a different model: that the bleeding advantage of a factor Xa inhibitor over a vitamin K antagonist doesn’t depend on renal clearance so much as it depends on the drug’s structural safety margin - narrower fluctuation, no interaction with dietary vitamin K, no INR volatility. Warfarin’s problems in CKD 4 don’t necessarily get smaller as the kidneys fail further. This is the study that started moving apixaban in advanced CKD from “avoid” toward “actively consider,” and it’s worth sitting with, because everything that follows in stage 5 is an attempt to answer whether that logic holds once dialysis enters the picture.

The Silence at Stage 5

Before getting to dialysis, it’s worth naming the gap that nobody likes to talk about: CKD stage 5, pre-dialysis, CrCl under 15 but not yet on renal replacement therapy. There is no dedicated randomized trial here. None. It’s the single largest hole in this entire evidence map, and it means that for a meaningful population of patients, clinicians are extrapolating from either the CKD 4 subgroup above or the dialysis trials below - neither of which quite describes their patient.

Subscribe now

Three Trials, One Consistent Whisper

Once a patient starts hemodialysis, we do finally have randomized data. Three trials, in fact. None of them individually proves much on its own, and that’s worth being honest about rather than glossing over. What makes them worth discussing together is that they keep landing in the same place despite being conducted in different countries, with different comparator regimens, and different primary endpoints.

RENAL-AF was supposed to be the trial that settled this - apixaban 5 mg twice daily, the standard dose with reduction to 2.5 mg for patients meeting age or weight criteria, against warfarin, in dialysis patients with CHA₂DS₂-VASc ≥2, powered for 760 patients. It enrolled 154 before stopping for slow recruitment. What’s left is underpowered by design: major or clinically relevant non-major bleeding occurred in 26% of the apixaban group versus 22% on warfarin, HR 1.20 (0.63–2.30) - a confidence interval wide enough to be compatible with either meaningful harm or meaningful benefit, which tells you almost nothing on its own. What RENAL-AF did contribute, and this matters more than the headline result, was pharmacokinetic data showing that apixaban 5 mg BID in dialysis patients produces drug levels comparable to what you’d see in a patient with CrCl 30–59. That single finding turns out to be more useful for the dosing debate later in this piece than the trial’s underpowered efficacy comparison.

AXADIA-AFNET 8 ran the opposite dose strategy - apixaban 2.5 mg twice daily against phenprocoumon, in 97 German dialysis patients followed for a median of roughly 14 months, the longest follow-up of any head-to-head apixaban-versus-VKA trial in this population. The composite safety endpoint occurred in 45.8% of the apixaban arm versus 51.0% on VKA, HR 0.93 (0.52–1.65). Non-inferiority was not established (p=0.16), and taken alone that reads as a negative trial. But look at where the point estimate sits: still favoring apixaban, just not narrowly enough to declare victory with 97 patients. Combined with what RENAL-AF found about drug exposure, AXADIA raises a genuinely important possibility - that 2.5 mg twice daily may simply be an under-dosed regimen in this population, diluting whatever benefit the drug class actually offers.

Valkyrie asked a different question and, despite being the smallest of the three in some respects, gave the most striking answer. De Vriese and colleagues randomized 132 dialysis patients to VKA, rivaroxaban 10 mg daily, or rivaroxaban plus vitamin K2, originally to study vascular calcification, with cardiovascular events as an endpoint. In the extended follow-up (median 1.88 years - longer than either of the apixaban trials above, though it’s answering a different question with a different drug), major cardiovascular events occurred at a rate of 63.8 per 100 patient-years with VKA, compared to 26.2 with rivaroxaban and 21.4 with rivaroxaban plus K2. Life-threatening or major bleeding was significantly lower with rivaroxaban than VKA - HR 0.39 (0.17–0.90), p=0.03. This is a small, single-purpose trial that wasn’t built to be definitive, and it shouldn’t be treated as such. But a threefold difference in cardiovascular events, alongside a statistically significant halving of major bleeding, is not the kind of signal you file away and forget. It’s the kind that justifies the larger trials now underway.

None of these three studies would survive scrutiny as a standalone basis for changing practice. Together, though, they describe a pattern that’s hard to dismiss as noise: three different DOAC-based strategies, three different countries, three trials too small to reach their own primary endpoints - and not one of them pointed toward warfarin being the safer choice.

What Real-World Data Told Us That Trials Couldn’t

Randomized trials in dialysis populations will always be small, because recruitment is genuinely hard in patients undergoing three-times-weekly hemodialysis with multiple competing comorbidities. This is exactly the kind of gap that a well-matched observational cohort can partially fill, and Siontis and colleagues did that using USRDS-linked Medicare claims from 2010 to 2015 - 25,523 patients, 2,351 on apixaban, propensity-matched 1:3 against warfarin.

The scale changes what you can see. Major bleeding was reduced with apixaban, HR 0.72 (0.59–0.87) - a tighter, more convincing interval than any of the three RCTs could produce individually. But the detail that matters most for practice sits in how the drug was dosed. Patients on the full 5 mg twice-daily dose had lower rates of both stroke (HR 0.64) and death (HR 0.63) compared to warfarin. Patients on the reduced 2.5 mg dose showed no thromboembolic advantage over warfarin at all.

Sit with that contrast for a moment, because it’s not a subtle finding. In this cohort, the dose most commonly prescribed by default, 2.5 mg, driven mechanically by dialysis status meeting one of the FDA’s dose-reduction criteria, was the dose that failed to outperform warfarin on stroke prevention. The dose that many prescribers avoid out of caution, 5 mg twice daily, was the one associated with better outcomes on every axis measured.

The Dose Nobody Quite Agrees On

Put the pieces from RENAL-AF, AXADIA, and Siontis next to each other and a genuinely uncomfortable question emerges: is the standard practice of dose-reducing apixaban to 2.5 mg in dialysis patients, largely a labeling artifact of age, weight, and creatinine criteria never validated specifically for dialysis pharmacokinetics, actually the more conservative choice, or is it quietly under-treating a population that already carries elevated stroke risk?

The pharmacokinetic data offers a plausible mechanism rather than just a correlation. RENAL-AF showed that 5 mg twice daily in dialysis patients produces exposure roughly equivalent to a patient with CrCl 30–59 - squarely within the range where DOAC efficacy is well established. If that’s true, then reflexively cutting the dose in half for dialysis patients may be solving a problem, supratherapeutic exposure, that doesn’t actually exist at this degree of renal replacement, while creating a different one: a dose too low to prevent the strokes it’s meant to prevent. AXADIA’s failure to reach non-inferiority at 2.5 mg, and Siontis’s finding that only the 5 mg dose showed benefit, both point the same direction.

None of this is settled enough to overturn labeling on its own. But it reframes the clinical conversation. The question in front of a physician managing a dialysis patient with AF isn’t simply “DOAC or warfarin.” It’s increasingly “which dose of which DOAC,” and the honest answer right now is that the dose most people are prescribing by default may not be the one the pharmacokinetic and outcomes data actually support.

Subscribe now

Where the Evidence Runs Out - For Now

KDIGO’s Controversies Conferences, in both 2018 and 2024, have been unusually direct about naming this as one of nephrology and cardiology’s most pressing unresolved questions, and the field has responded with several trials now underway, SAFE-D, AVKDIAL, and ISCHEMIA-CKD among them, designed specifically to give CKD stage 4 through dialysis-dependent patients the kind of adequately powered evidence that RENAL-AF, AXADIA, and Valkyrie couldn’t individually deliver. There’s also a growing conversation about whether pharmacologic anticoagulation is even the right frame for the highest-risk end of this population, with left atrial appendage occlusion increasingly discussed as a non-pharmacologic alternative for patients whose bleeding risk makes any anticoagulant, at any dose, a genuinely difficult trade.

What strikes me most, working back through this evidence, isn’t that we lack answers in advanced CKD and dialysis. It’s that the answers we do have arrived through a strange collective effort - three trials that individually failed to reach their targets, one observational cohort large enough to catch what the trials couldn’t, and a pharmacokinetic sub-study that quietly reframes how we think about dosing. No single piece is definitive. The pattern across all of them is harder to ignore than any one result would be alone.

The question worth carrying forward isn’t whether we should anticoagulate dialysis patients with atrial fibrillation - that decision, as a general principle, isn’t really in doubt, even as the LAAO conversation above makes clear which patients fall outside it. It’s whether the dosing conventions we inherited from a labeling process built around age, weight, and a creatinine clearance number that stops meaning much once someone starts dialysis, are actually serving the patients they were designed to protect.

To accompany this article, I've made the complete one-page Cardio Concepts framework available as a free download, summarizing the complex evidence on anticoagulation across the spectrum of advanced CKD and dialysis into a single visual reference.

The Renal Cliff

Previous When Every Drug Makes Something Worse: Anticoagulation in AF Patients After PCI

Next The 120-Kilogram Line We Drew From Silence, Not Data

References

1. Pokorney SD, Chertow GM, Al-Khalidi HR, et al. Apixaban for Patients With Atrial Fibrillation on Hemodialysis: A Multicenter Randomized Controlled Trial. Circulation. 2022;146:1735–1745. DOI: 10.1161/CIRCULATIONAHA.121.054990

2. Reinecke H, Engelbertz C, Bauersachs R, et al. A Randomized Controlled Trial Comparing Apixaban With the Vitamin K Antagonist Phenprocoumon in Patients on Chronic Hemodialysis: The AXADIA-AFNET 8 Study. Circulation. 2023;147:296–309. DOI: 10.1161/CIRCULATIONAHA.122.062779

3. De Vriese AS, Caluwé R, Pyfferoen L, et al. Multicenter Randomized Controlled Trial of Vitamin K Antagonist Replacement by Rivaroxaban with or without Vitamin K2 in Hemodialysis Patients with Atrial Fibrillation: the Valkyrie Study. Journal of the American Society of Nephrology. 2020;31:186–196. DOI: 10.1681/ASN.2019060579 (extended follow-up: JASN. 2021;32:1474–1483.)

4. Siontis KC, Zhang X, Eckard A, et al. Outcomes Associated With Apixaban Use in Patients With End-Stage Kidney Disease and Atrial Fibrillation in the United States. Circulation. 2018;138:1519–1529. DOI: 10.1161/CIRCULATIONAHA.118.035418

5. Stanifer JW, Pokorney SD, Chertow GM, et al. Apixaban Versus Warfarin in Patients With Atrial Fibrillation and Advanced Chronic Kidney Disease. Circulation. 2020;141:1384–1392. DOI: 10.1161/CIRCULATIONAHA.119.044059

   Share

   Leave a comment

   Explore the Cardio Concepts Library

There’s a particular kind of clinical discomfort that comes with managing a patient who has atrial fibrillation and just had a coronary stent placed. You need anticoagulation for the AF - the stroke risk is real and measurable. You need antiplatelet therapy for the stent - skip it and you’re risking acute thrombosis. Stack them together and you’ve created a bleeding hazard that’s hard to ignore.

For years, the default answer was triple therapy: warfarin, aspirin, and a P2Y12 inhibitor. It made intuitive sense - each component had a clear rationale, each addressed a different mechanism, and for a long time nobody seriously questioned whether the combination itself was the problem.

That changed around 2013, starting with a trial most people initially dismissed as too small to matter.

The First Crack: Removing Aspirin

WOEST was 573 patients, open-label, mostly European centers. It asked a question that felt almost heretical at the time: what happens if you just drop the aspirin?

Patients on dual therapy (warfarin plus clopidogrel, no aspirin) had a bleeding rate of 19.4% versus 44.4% in the triple therapy group over one year. HR 0.36, NNT roughly 4. And critically, ischemic events didn’t increase - both total ischemic events and all-cause mortality (2.5% vs 6.3%) came out nominally better in the dual therapy arm, though neither finding should be taken at face value from a trial this size and not designed to detect those differences.¹

The pushback was predictable: underpowered, not designed to detect ischemic differences, mixed AF and non-AF population. All fair. But WOEST forced a question that hadn’t been seriously examined before - what is aspirin actually contributing here?

What the trial data made increasingly hard to ignore was this: with a P2Y12 inhibitor already suppressing ADP-mediated platelet aggregation, aspirin’s incremental antithrombotic contribution in this specific setting appeared modest at best. The bleeding cost, as WOEST showed, was anything but modest.

Expanding the Question Into the NOAC Era

WOEST still used warfarin. The next question was whether replacing the OAC component with a NOAC would change the picture further.

PIONEER AF-PCI (2016) brought rivaroxaban into this space for the first time - three arms: rivaroxaban 15 mg daily plus a P2Y12 inhibitor, rivaroxaban 2.5 mg twice daily (a dose lower than standard AF dosing) plus DAPT, or VKA-based triple therapy. Clinically significant bleeding: 16.8%, 18.0%, 26.7% respectively. Both rivaroxaban arms outperformed triple therapy on safety.²

PIONEER wasn’t powered for ischemic outcomes, and that was a real limitation. But it demonstrated that an OAC-plus-single-antiplatelet approach was viable in a real trial population - enough to make the next generation of trials ethically defensible.

RE-DUAL PCI (2017) tested dabigatran at two doses - 110 mg and 150 mg twice daily, each paired with a P2Y12 inhibitor - against triple therapy. Dabigatran 110 mg reduced major or clinically relevant non-major bleeding from 26.9% to 15.4% (HR 0.52). The 150 mg dose also outperformed triple therapy, with a somewhat smaller margin. Non-inferiority on ischemic outcomes was achieved for both doses combined.³

One finding worth sitting with: the 110 mg dose showed a numerical signal toward increased stent thrombosis compared to 150 mg, though it didn’t reach significance. It doesn’t change the overall picture, but it stays relevant when thinking about dose selection in post-ACS patients where early ischemic risk is highest.

The Cleanest Answer: AUGUSTUS

AUGUSTUS (2019) is where the evidence became most actionable. The 2×2 factorial design enrolled 4,614 patients with AF and recent ACS or PCI, testing two independent decisions simultaneously: apixaban versus VKA, and aspirin versus placebo, both on top of a P2Y12 backbone - with a follow-up of six months.

Earlier trials had compared bundles. AUGUSTUS compared components - which is exactly what clinicians actually need.

Apixaban versus VKA: 10.5% versus 14.7% major or clinically relevant non-major bleeding (HR 0.69, NNT 24), with fewer hospitalizations and deaths in the apixaban arm.

Aspirin versus placebo: 16.1% versus 9.0% bleeding (HR 1.89, NNH 14), with no detectable improvement in ischemic outcomes.

The range across the four combinations makes it concrete. Apixaban plus P2Y12, no aspirin: 7.3% bleeding. VKA plus DAPT: 18.7%.⁴ More than a doubling of bleeding risk, driven by two choices that are entirely within the prescriber’s control.

Subscribe now

What the Synthesis Tells Us

The network meta-analysis from Lopes and colleagues pooled five RCTs - WOEST, PIONEER, RE-DUAL PCI, AUGUSTUS, ENTRUST-AF PCI - roughly 11,500 patients total. NOAC plus single P2Y12 inhibitor came out with the lowest bleeding burden (OR 0.52 for major TIMI bleeding versus VKA plus DAPT), intracranial hemorrhage reduced by about two-thirds.⁵

One signal that keeps appearing across subgroup analyses and shouldn’t be ignored: in ACS patients, removing aspirin entirely from the early post-procedure window may increase short-term stent thrombosis risk in the first 30 days. The evidence isn’t definitive, but it’s consistent enough that most guidelines treat ACS differently from stable PCI - recommending a brief aspirin-inclusive period before transitioning to dual therapy.

This is exactly where a blanket protocol breaks down. A complex bifurcation stent in the context of an anterior STEMI isn’t the same ischemic situation as elective PCI for a single-vessel stable lesion. The framework is dual therapy as the default; the judgment is how much, if any, aspirin overlap that specific patient warrants.

Thinking About What Changed

Triple therapy persisted for as long as it did partly because the question was never asked. Three drugs, three rationales - there was no obvious candidate to remove. Aspirin had been part of coronary disease management for decades, and removing it felt like dismantling something foundational. That instinct was never tested in this specific population. Nobody formally examined whether the foundational assumption held when a P2Y12 inhibitor was already in the picture.

The other part is how bleeding was framed. In the triple therapy era, bleeding wasn’t seen as a problem created by the regimen - it was treated as an inherent cost of treating these patients. The clinical response was gastroprotection, careful INR management, close monitoring. The possibility that the bleeding was partly self-inflicted by the aspirin didn’t enter the conversation until WOEST made it unavoidable.

What the subsequent trials confirmed, one by one, is that the relevant question was never “how do we tolerate triple therapy?” It was whether the third drug was doing enough to justify being there. For most patients, it wasn’t.

Where We Still Don’t Have Answers

Despite how quickly this field moved in the 2013–2020 period, some gaps remain meaningful.

No head-to-head RCTs compare NOACs directly in this population. Apixaban has become the practical default for many clinicians - reasonable given AUGUSTUS and its overall profile - but that reflects inference, not direct comparative evidence.

The P2Y12 choice is largely unaddressed by trial data. More than 90% of patients in the major AF-PCI trials received clopidogrel. In real-world ACS management, ticagrelor and prasugrel are used routinely. Their stronger antiplatelet effect almost certainly shifts the bleeding calculus, but how much is still hypothesis.

And the timing of aspirin de-escalation remains a judgment call. How long is “brief”? Discharge, one week, one month? No dedicated trial has answered this with precision. Most clinicians set their own threshold based on their read of the ischemic risk - which is reasonable, but means the decision is being made without much direct evidence underneath it.

The Practical Takeaway

For most AF patients undergoing PCI today, dual therapy is the starting point: a NOAC plus a P2Y12 inhibitor. Aspirin gets added when the early ischemic risk - ACS presentation, complex stent anatomy, high residual disease burden - seems to warrant it, and only for a defined short window. Defining that window means weighing bleeding risk against ischemic risk for that specific patient, using whatever tools are available: HAS-BLED, DAPT score, PRECISE-DAPT, clinical judgment.

But the default has genuinely shifted. Aspirin now requires a reason to be included, not a reason to be removed. That’s the practical consequence of a decade of trials - not a new drug or a new mechanism, just a more honest accounting of what each component in the regimen is actually doing.

To accompany this article, I've made the complete one-page Cardio Concepts framework available as a free download, bringing the evidence behind modern AF-PCI antithrombotic therapy into a single visual reference.

Burden of Proof

—Cardio Concepts

Previous When the Gut Bleeds and the Heart Still Needs Protecting

Next The Renal Cliff: Why Our Confidence in Anticoagulation Collapses Exactly Where Patients Need It Most

References

1. Dewilde WJM, et al. WOEST. Lancet. 2013;381(9872):1107–1115. doi:10.1016/S0140-6736(12)62177-1

2. Gibson CM, et al. PIONEER AF-PCI. N Engl J Med. 2016;375(25):2423–2434. doi:10.1056/NEJMoa1611594

3. Cannon CP, et al. RE-DUAL PCI. N Engl J Med. 2017;377(16):1513–1524. doi:10.1056/NEJMoa1708454

4. Lopes RD, et al. AUGUSTUS. N Engl J Med. 2019;380(16):1509–1524. doi:10.1056/NEJMoa1817083

5. Lopes RD, et al. JAMA Cardiol. 2020;5(5):582-589. doi:10.1001/jamacardio.2019.6175

Share

Leave a comment

Explore the Cardio Concepts Library

There’s a particular kind of clinical discomfort that comes with a patient who just had a significant GI bleed, actively or recently, and also carries AF with a CHA₂DS₂-VASc of 4 or 5. The gastroenterologist has done their job. The scope is clean, the bleeder is clipped, hemostasis is secure. Now everyone looks at you.

Do you restart anticoagulation? When? With what?

The reflex answer for a long time was: not yet, maybe never, let’s see. And honestly, that answer came from somewhere reasonable, the fear of re-bleeding felt immediate and concrete. The stroke felt like something that might happen later, somewhere else. But that’s exactly the kind of asymmetry that has quietly led to a lot of preventable strokes.

The Question Witt Asked First

Back in 2012, Witt and colleagues did something surprisingly simple that hadn’t really been done before: they followed 442 warfarin patients who had a GI bleed, and asked what actually happened over the next 90 days depending on whether anticoagulation was restarted or not.

The numbers were stark. Patients who restarted warfarin had a thromboembolism rate with an HR of 0.05 compared to those who didn’t - essentially an order-of-magnitude reduction. Mortality HR was 0.31. The re-bleeding risk? HR 1.32, not statistically significant.

This was a small retrospective cohort from a single center, and nobody should mistake it for definitive evidence. But it framed a question that the field had been avoiding: not whether to restart anticoagulation, but what the actual trade-off looks like when you put numbers on both sides. And the numbers, even in this limited dataset, were not subtle.

Scaling Up the Observation

The Danish registry data that Staerk and colleagues published in 2015 took the same question and answered it at a scale that made the signal much harder to ignore - over 4,600 AF patients after a major GI bleed, drawn from 16 years of national registry data, with a median follow-up of 2 years per patient.

Patients who restarted anticoagulation as monotherapy had mortality HR 0.39 and thromboembolic risk HR 0.41. Re-bleeding risk did increase - HR 1.37 - but the net balance tilted clearly toward restarting.

One finding worth sitting with: antiplatelet therapy alone didn’t help. Patients who received aspirin or clopidogrel as a substitute for anticoagulation saw no meaningful reduction in stroke or death. The “let’s just use aspirin for now” approach - still common as a hedge - doesn’t actually reduce the risk you’re most worried about. It just feels like you’re doing something.

The question of when to restart doesn’t have a clean RCT answer. Most guidance has settled around roughly 7 to 30 days after hemostasis - early enough to limit thromboembolic exposure, late enough that re-bleeding risk from the index event has settled. But this isn’t a hard number from a controlled trial. It’s a working estimate synthesized from observational data, and it’s worth keeping that distinction in mind when the patient in front of you doesn’t fit neatly into the window.

A New Problem: The DOACs Aren’t All the Same

Once the major DOAC trials came in, a different question emerged. Not just whether to restart, but which drug - and whether the assumption that “DOACs are generally safer on the GI tract” held up when you looked at specific agents.

The ROCKET-AF GI substudy (Sherwood et al., JACC 2015) was clarifying in an uncomfortable way. Rivaroxaban at standard dose had significantly higher rates of major or clinically relevant GI bleeding compared to warfarin - 3.61% vs 2.60% per year, HR 1.42. Not a subtle signal. It also reinforced something worth building into how you think about these patients: prior GI bleed history was an independent predictor of future GI bleeding on treatment. A patient who’s bled before is telling you something about their underlying vulnerability. That history doesn’t disappear after the scope report is filed.

But ROCKET-AF only compared rivaroxaban to warfarin, in one trial, in one population. It couldn’t tell you how the four available OACs compared to each other - which is the actual clinical question when you’re trying to pick the right drug for a specific patient.

The Hierarchy No One Was Prepared For

The Ray et al. study in JAMA 2018 is where the picture got more specific, and more actionable. Using Medicare claims data on over 1.6 million new OAC users, they compared upper GI bleeding hospitalization rates across all four major OACs. What came out:

Rivaroxaban 144 → Dabigatran 120 → Warfarin 113 → Apixaban 73 per 10,000 patient-years.

Apixaban came in below warfarin. Not just below the other DOACs - below warfarin. IRR for apixaban versus warfarin was 0.64. For rivaroxaban versus apixaban: 1.97.

The class-level narrative had been that all DOACs carried broadly similar GI profiles, with maybe some variation at the edges. This data said something different: within the class, you’re not choosing between similar options. The gap between rivaroxaban and apixaban on upper GI bleeding is large enough that it should factor into your decision - especially in a patient who’s already shown they can bleed from the gut.

There are likely pharmacological explanations for these differences, though the precise mechanisms remain incompletely characterized. What matters clinically is less the mechanism than what the data means for the patient in front of you.

The PPI Finding: Simple, Underused

Ray et al. also quantified something that had been assumed but not confirmed at this scale: PPI co-therapy reduces upper GI bleeding hospitalization by roughly 34% across OAC users, overall IRR 0.66.

The protective effect was largest with dabigatran - IRR fell to 0.49. This makes pharmacological sense. Dabigatran’s formulation contains tartaric acid within the pellet core, which lowers local pH to facilitate drug dissolution and absorption. The trade-off is that the gastric mucosa ends up exposed to a more acidic local environment, making it more susceptible to acid-related injury. PPI blunts that exposure.

Among high-risk patients - those with prior GI bleed, concurrent NSAID use, or advanced age - the NNT for PPI co-therapy was 113. That figure becomes more meaningful when you weigh it against how cheap, low-risk, and easy to prescribe a PPI actually is. For these patients, adding a PPI isn’t a marginal consideration. It should be part of the restart plan.

Putting This Together at the Bedside

The clinical picture that comes out of all this isn’t really a protocol - it’s more like a set of commitments that require active reasoning rather than a default.

The first is that restarting anticoagulation in a patient with significant thromboembolic risk is not really optional for most patients. The survival data from Witt and Staerk are consistent enough that “hold everything” isn’t actually the conservative option. It has a cost - just one that’s less immediately visible.

The timing question comes down to making sure the bleeding source has been definitively managed, not just controlled. That means the endoscopist has done what needs doing before you think about restarting the anticoagulant. The 7–30 day window is a reasonable working frame, but the clinical readiness of the patient matters more than hitting a calendar target.

Drug selection is where the Ray data become practically useful. In a patient with prior GI bleeding who needs an OAC, there’s a defensible case for preferring apixaban - lower upper GI bleeding rates than warfarin, and substantially lower than rivaroxaban. Not because a guideline mandates it, but because the evidence is coherent and the difference is clinically meaningful.

And then add the PPI if it isn’t already there.

What We Still Don’t Know

There’s no RCT comparing OACs head-to-head specifically in patients with prior GI bleeding. The Witt and Staerk data are observational. ROCKET-AF’s GI substudy tells you about one drug versus warfarin in a trial population. Ray et al. is retrospective claims data - large-scale and directionally informative, but not randomized. The guidance that comes out of this evidence base is better than nothing, but it’s built on a foundation that has real limits.

Lower GI bleeding - small bowel, right colon - is also largely unstudied in this context. The pharmacological predictors may differ from upper GI, and the post-bleed management considerations aren’t identical. Applying the same framework may not be appropriate.

For patients with recurrent or severe GI bleeding where the re-bleed risk feels too high for any oral anticoagulant, left atrial appendage occlusion is increasingly worth raising earlier than it used to be. The evidence in this specific indication is still mostly observational, but it’s no longer a last-resort conversation.

Subscribe now

A Final Thought

When a patient has bled, re-bleeding risk tends to weigh heavier than stroke risk in how clinicians make decisions - not because the numbers justify it, but because of how we experience accountability. A stroke that occurs while the patient is off anticoagulation registers differently than a re-bleed after you restarted. One feels like the disease; the other feels like a choice you made.

That’s a real pull, and it’s worth being honest about it. Because Witt and Staerk both show that the mortality benefit of restarting is substantial, while the re-bleed risk increase - real as it is - is statistically modest and partially modifiable through drug selection and GI protection. A patient who’s been sitting off anticoagulation for six weeks while everyone waits for the “right moment” isn’t in a holding pattern. They’re accumulating stroke risk with nothing protecting them.

To accompany this article, I've made the complete one-page Cardio Concepts framework available as a free download for your clinical reference.

The Restart Calculus

—Cardio Concepts

Previous When Falling Doesn't Mean You Should Stop the Anticoagulant

Next When Every Drug Makes Something Worse: Anticoagulation in AF Patients After PCI

References:

1. Witt DM, Delate T, Garcia DA, et al. Risk of thromboembolism, recurrent hemorrhage, and death after warfarin therapy interruption for gastrointestinal tract bleeding. Arch Intern Med. 2012;172(19):1484–1491. doi:10.1001/archinternmed.2012.4261

2. Staerk L, Lip GYH, Olesen JB, et al. Stroke and recurrent haemorrhage associated with antithrombotic treatment after gastrointestinal bleeding in patients with atrial fibrillation. BMJ. 2015;351:h5876. doi:10.1136/bmj.h5876

3. Sherwood MW, Nessel CC, Hellkamp AS, et al. Gastrointestinal bleeding in patients with atrial fibrillation treated with rivaroxaban or warfarin: ROCKET AF trial. J Am Coll Cardiol. 2015;66(21):2271–2281. doi:10.1016/j.jacc.2015.09.024

4. Ray WA, Chung CP, Murray KT, et al. Association of oral anticoagulants and proton pump inhibitor cotherapy with hospitalization for upper gastrointestinal tract bleeding. JAMA. 2018;320(21):2221–2230. doi:10.1001/jama.2018.17242

Share

Leave a comment

Explore the Cardio Concepts Library

There’s a clinical reflex that most of us have seen - or done ourselves. An elderly patient with atrial fibrillation comes in, the chart mentions frequent falls, and somewhere in the reasoning chain, anticoagulation quietly gets deprioritized. Maybe it gets stopped altogether. The logic feels intuitive: this person falls, bleeding risk is high, let’s not make things worse.

The problem is that this reflex has probably caused more strokes than the evidence would justify.

For a long time, aspirin was the default when clinicians felt warfarin was “too risky” in elderly patients with AF. The reasoning was never particularly evidence-based, but it felt safer. Aspirin was familiar, didn’t require INR monitoring, and didn’t carry the same psychological weight as “blood thinners.” It was a compromise that felt defensible - even when it probably wasn’t.

BAFTA changed that in 2007. The trial enrolled 973 patients with AF aged 75 and older - average age 81.5 - across 260 primary care practices in England and Wales. These weren’t carefully selected patients. They were the real elderly population sitting in family medicine waiting rooms. Warfarin targeting INR 2.0–3.0 was compared head-to-head against aspirin 75 mg, and the result was unambiguous: the primary composite of fatal or disabling stroke, intracranial haemorrhage, and arterial embolism occurred at 1.8% per year on warfarin versus 3.8% per year on aspirin - a relative risk reduction of 52%, NNT roughly 50 per year. Major bleeding rates: 1.9% versus 2.2%, statistically indistinguishable.

The trial didn’t just show warfarin was better. It showed that the assumption underlying the aspirin compromise - that aspirin was meaningfully safer - was simply wrong. For an 81-year-old with AF, aspirin offered the bleeding risk of anticoagulation without anything close to the protection.

The concern about falls and anticoagulation has always been specifically about intracranial hemorrhage after head trauma. An elderly patient on warfarin who strikes their head after a fall faces a genuinely different risk profile than a younger patient. This is not an unreasonable worry - subdural hematomas in anticoagulated elderly patients are serious, and some are fatal.

What the worry tends to miss is the math.

Man-Son-Hing and colleagues worked through the arithmetic in a decision analysis model that has been widely cited in geriatric anticoagulation practice: to neutralize the stroke-prevention benefit of warfarin in a high-risk AF patient, that patient would need to fall approximately 295 times per year. The benefit of prevented strokes is so large, and fall-related intracranial hemorrhage so relatively infrequent, that the threshold is essentially unreachable in any real clinical scenario.

Donzé and colleagues put this to empirical test in 2012. Following 515 VKA-anticoagulated patients at a Swiss center for 12 months - nearly 60% classified as high fall risk - they found no significant difference in major bleeding between high and low fall-risk groups: 8.0 versus 6.8 events per 100 patient-years, HR 1.09, p = 0.64. Fall-related bleeds specifically accounted for just 0.6 per 100 patient-years. The study was single-center and VKA-only, which limits how far you can push the findings, but the signal was consistent with everything else in the literature.

What both pieces of evidence pushed toward was a different question. Not “is this patient safe to anticoagulate?” but “what does this specific patient actually gain or lose?” - and those two questions don’t always have the same answer.

The arrival of DOACs added another layer. Not just because they’re easier to manage than warfarin, but because they changed the risk arithmetic in a way that’s particularly relevant for older patients.

The falls subanalysis of ENGAGE AF-TIMI 48, published by Steffel and colleagues in JACC 2016, illustrates this. Among 21,000+ patients in the trial, about 900 were identified as high fall risk. This group had worse outcomes across the board - higher major bleeding, more fractures, higher mortality - which makes sense, because fall risk in this population is largely a proxy for overall frailty. The treatment interaction was the telling part: edoxaban performed just as well relative to warfarin in the high fall-risk group as in everyone else. And because this group had higher baseline event rates, the absolute benefit of edoxaban over warfarin ended up numerically larger in fall-risk patients than in the overall trial population.

It’s not a paradox once you think it through. Patients with more to lose from a stroke also have more to gain from effective prevention. Fall risk doesn’t change that equation - it just makes the stakes on both sides higher.

Still, there’s a phenotype that most landmark trials didn’t enroll in meaningful numbers: the very elderly - mid-to-late eighties, low body weight, impaired renal function, prior bleeding history. Where standard-dose DOAC isn’t appropriate, and where the evidence has historically been thinnest.

ELDERCARE-AF addressed this directly. The trial enrolled 984 patients in Japan aged 80 and older - average age 86.6 - who had AF but were ineligible for standard-dose anticoagulation due to at least one of: creatinine clearance 15–30 mL/min, body weight ≤45 kg, or prior moderate-to-severe bleeding. These were patients many clinicians would have left untreated, or managed with aspirin by default.

They were randomized to edoxaban 15 mg daily - half the already-reduced 30 mg dose - versus placebo. Stroke reduction: 2.3% per year versus 6.7% per year, a 66% relative risk reduction. Major bleeding increased numerically (3.3% vs 1.8%), but the p-value was 0.09 and there was no increase in intracranial hemorrhage.

“Not suitable for standard dosing” doesn’t automatically mean “not suitable for anticoagulation.” ELDERCARE-AF opened a clinical space that previously didn’t have RCT support - though it also leaves an unanswered question about whether the same approach would work with apixaban 2.5 mg BID, which is used widely in practice outside East Asia without an equivalent trial in this population.

FRAIL-AF, published in Circulation 2024, came with a result that most people didn’t see coming.

The trial enrolled 1,330 patients aged 75 and older with AF and meaningful frailty - Groningen Frailty Indicator ≥3, average age 83. All were already on VKA therapy. The question was pragmatic: should we switch these patients to DOAC? Less monitoring, fewer food interactions, lower intracranial hemorrhage rates in the general AF population - the assumption was that switching would be a straightforward improvement.

Major bleeding and clinically relevant non-major bleeding: 15.3% in the DOAC group versus 9.4% in the VKA group, HR 1.69. The trial stopped early for futility. Thrombotic events were equivalent.

This doesn’t indict DOACs in elderly patients broadly. The finding is more specific: in patients who are already stable on VKA - titrated, monitored, tolerating it - switching introduced net harm. Frail patients have more variable drug absorption and renal clearance, which produces less predictable anticoagulant exposure regardless of which DOAC is chosen. A patient maintaining stable INR on warfarin has, in a way, already been individually optimized for their own pharmacokinetics. Switching disrupts that without a guaranteed gain.

The harder lesson is about how we apply evidence. “DOACs are better than warfarin” holds across the general AF population - but a frail patient who is stable on VKA is a different clinical situation from a frail patient initiating anticoagulation for the first time. FRAIL-AF is a reminder that population-level evidence doesn’t always transfer cleanly to the individual in front of you.

Subscribe now

Two decades of evidence have sharpened the contours of this problem without fully resolving it.

Falls alone are not a reason to withhold anticoagulation - the quantitative case against that reflex is now hard to argue. For patients initiating therapy, DOACs are generally the better choice, and the absolute benefit is often largest precisely in the higher-risk patients we worry most about. For those who are very elderly and don’t qualify for standard dosing, ELDERCARE-AF gives us an RCT-supported option we didn’t have before. But for frail patients already stable on VKA, the routine switch to DOAC deserves real pause. FRAIL-AF doesn’t say never switch - it says don’t assume the switch is an improvement.

What’s still missing is a reliable way to make these distinctions at the bedside. Which frail patient benefits from DOAC initiation? Which one should stay on VKA? Who might be better served by left atrial appendage occlusion? Comprehensive geriatric assessment and frailty-integrated risk tools are being studied, but none have made it into routine clinical workflow yet.

The patients who most need anticoagulation are often the hardest to anticoagulate safely. That gap hasn’t closed. But at least now we know more precisely where it sits - and where our assumptions have been quietly working against the evidence.

To make the evidence easier to use in daily practice, I've turned the key trials into a one-page framework. It's free to download and share with your team

THE FALLS FALLACY

—Cardio Concepts

Previous Beyond Stroke Prevention

Next When the Gut Bleeds and the Heart Still Needs Protecting

Subscribe now

References

1. Mant J, et al. Warfarin versus aspirin for stroke prevention in an elderly community population with atrial fibrillation (BAFTA). Lancet. 2007;370(9586):493–503. doi:10.1016/S0140-6736(07)61233-1

2. Man-Son-Hing M, Laupacis A. Anticoagulant-related bleeding in older persons with atrial fibrillation: physicians’ fears often unfounded. Arch Intern Med. 2003;163(13):1580–1586. doi:10.1001/archinte.163.13.1580

3. Donzé J, et al. Risk of falls and major bleeds in patients on oral anticoagulation therapy. Am J Med. 2012;125(8):773–778. doi:10.1016/j.amjmed.2012.01.033

4. Steffel J, et al. Edoxaban versus warfarin in atrial fibrillation patients at risk of falling: ENGAGE AF–TIMI 48 Analysis. J Am Coll Cardiol. 2016;68(11):1169–1178. doi:10.1016/j.jacc.2016.06.034

5. Okumura K, et al. Low-dose edoxaban in very elderly patients with atrial fibrillation (ELDERCARE-AF). N Engl J Med. 2020;383(18):1735–1745. doi:10.1056/NEJMoa2012883

6. Joosten LPT, et al. Safety of switching from a vitamin K antagonist to a non–vitamin K antagonist oral anticoagulant in frail older patients with atrial fibrillation (FRAIL-AF). Circulation. 2024;149(4):279–289. doi:10.1161/CIRCULATIONAHA.123.066485

   Share

   Leave a comment

   Explore the Cardio Concepts Library

The ones who are: older, frailer, post-GI bleed, just got a stent, creatinine that moves every few months, weighs 140 kg, or had a stroke while already on anticoagulation. The evidence for these patients is messier, more conditional, and sometimes pulls in opposite directions. Moving from VKA to DOAC has reshaped how we approach almost all of them - but it hasn’t made the decisions cleaner. In some ways it’s surfaced how much of our earlier reasoning was built on assumptions that don’t hold when you’re treating an actual person rather than a trial population.

This series works through seven of these high-risk phenotypes. For each, the goal isn’t to summarize what the guidelines say. It’s to trace how the evidence actually evolved - which trials changed the way we think, where the real clinical dilemmas still live, and what “individualized anticoagulation” actually means when you’re looking at a specific patient in a specific situation.

In this series:

Part 1 — When Falling Doesn’t Mean You Should Stop the Anticoagulant
Why “she falls a lot” is not a reason to withhold anticoagulation — and what FRAIL-AF taught us about when it actually might be.

Part 2 — When the Gut Bleeds and the Heart Still Needs Protecting
Restarting anticoagulation after a GI bleed is almost always the right call. The harder question is how, when, and with what.

Part 3 — When Every Drug Makes Something Worse: Anticoagulation in AF Patients After PCI
How four trials systematically dismantled triple therapy — and why the aspirin-free strategy is now the default.

Part 4 — The Renal Cliff: Why Our Confidence in Anticoagulation Collapses Exactly Where Patients Need It Most
From CKD stage 3 to dialysis, the quality of evidence drops off sharply — and at the dialysis end, we’re mostly working with underpowered trials and registry data.

Part 5 — The 120-Kilogram Line We Drew From Silence, Not Data
The shift from “avoid DOACs above BMI 40” to “apixaban works across all weight categories” — how one subgroup analysis changed five years of practice.

Part 6 — The Patient Who Strokes Anyway
Breakthrough stroke is its own clinical entity — with a distinct mechanism, a distinct recurrence risk, and no simple fix.

Part 7 —When the Tumor Writes Its Own Rules for Clotting
Cancer-associated thrombosis isn’t just VTE in a sicker patient. The biology is different, and so is the evidence for how to treat it.

Explore the Cardio Concepts Library

But the deeper I went into the literature, the more I realized that this seemingly dry concept quietly shapes how we interpret entire studies. What counts as a hospitalization? Which events are included - and which remain invisible? How many times does a patient have to deteriorate before the trial acknowledges it?

Behind those questions is a surprisingly human story about how medicine chooses to measure worsening heart failure.

There is a peculiar moment in the history of heart failure trials worth pausing on. Two SGLT2 inhibitor trials, enrolling similar patients, testing drugs from the same class, produced hazard ratios that were nearly identical - yet EMPEROR-Reduced recorded more heart failure hospitalizations than DAPA-HF. Not because empagliflozin was more effective than dapagliflozin, or less. But because the two trials were counting differently. One included urgent outpatient visits requiring intravenous therapy. The other did not.

That discrepancy - quiet, methodological, easy to miss in a journal club - is the entry point into one of the most underappreciated problems in cardiovascular medicine. We have spent three decades treating hospitalization as an objective fact. A patient either was admitted or wasn’t. The number goes into the analysis. The trial reads out positive or negative.

But hospitalization is not a fact. It is a construct. And how we build that construct - what we count, how many times we count it, and which patients we count it in - shapes what we conclude about our treatments in ways that matter more than most of us realize when we read a trial.

The First Problem: We Only Counted Once

The Cox proportional hazards model is elegant and widely understood. It also has a fundamental limitation that took the heart failure community too long to reckon with: it only captures the first event.

In a disease defined by its tendency to recur - repeated decompensations, each one eroding functional reserve a little further - this is not a minor technical footnote. It is a structural misfit between the statistical model and the biology being studied.

The numbers from PARADIGM-HF make this concrete. When Mogensen and colleagues reanalyzed the trial using a joint frailty model that accounted for recurrent events, they found that of 3,181 total composite events, only 2,031 - roughly 63.8% - were first events. More than one in three hospitalizations had simply vanished from the primary analysis. When those recurrent events were brought back into the picture, the HR for sacubitril/valsartan moved from 0.80 to 0.75.¹ The drug had been doing more than the published number showed.

CHARM-Preserved tells an equally instructive story, though one that requires careful handling. In the primary time-to-first analysis, candesartan produced an HR of 0.86 with a p-value of 0.050 - landing in the uncomfortable territory just at the edge of significance, widely read as neutral. When Rogers and colleagues reanalyzed the data using a negative binomial model for recurrent events, the rate ratio moved to 0.75 with p=0.003.² This was a post-hoc reanalysis, not pre-specified, and it says nothing definitive about candesartan in HFpEF. But it raises an uncomfortable question: how often has the statistical framework we chose to evaluate a drug been structurally incapable of detecting what was actually in the data?

In a disease that recurs by nature, time-to-first was always going to be a partial answer.

The Second Problem: We Were Counting in the Wrong Place

Something was shifting in clinical practice while the field was still debating recurrent event models. Outpatient heart failure management was improving - specialist clinics, closer follow-up, flexible diuretic titration - and patients who would previously have been admitted for decompensation were increasingly being managed with intravenous furosemide at the clinic and sent home the same day.

This created a situation nobody had fully anticipated. The better the system became at avoiding formal hospitalization, the more the primary endpoint of heart failure trials began to drift from actual clinical deterioration. A patient treated with IV diuretics in the emergency department and discharged without an overnight stay would not appear in the hospitalization count. But her prognosis, her trajectory, her burden of illness was not meaningfully different from the patient in the next bed who happened to be admitted.

Analysis of worsening heart failure events in the PARADIGM-HF outpatient population found that episodes managed entirely in the outpatient setting - escalation of oral diuretics, intravenous therapy at clinic or the emergency department - carried prognostic weight comparable to formal hospitalization.³ None of it appeared in the trial’s primary endpoint.

SOLOIST-WHF was the first trial to formally build this reality into its design. By including urgent visits requiring intravenous therapy alongside traditional hospitalization in the primary composite, the trial captured event rates of 51.0 versus 76.3 per 100 patient-years - an HR of 0.67 (95% CI 0.52–0.85) for sotagliflozin.⁴ The trial was terminated early for financial reasons related to the pandemic and enrolled only patients with diabetes, which limits what can be concluded about the drug. But as a demonstration that the definition of the endpoint changes what you measure - and therefore what you find - it was clarifying.

DELIVER then provided the more robust evidence. In 6,263 patients with HFpEF and HFmrEF, dapagliflozin reduced a composite that included both heart failure hospitalization and urgent heart failure visits. Approximately 13% of the nonfatal events in that trial were urgent visits - events that would have been invisible under a conventional endpoint definition.⁵

But even with the right endpoint in place, the right patients still have to be in the trial. TOPCAT is the cautionary example that deserves more discomfort than it usually gets.

In the overall analysis, spironolactone appeared neutral in HFpEF. When Pfeffer and colleagues examined the regional data, event rates in Russia and Georgia were roughly four times lower than in the Americas - a discrepancy too large to explain away by baseline differences.⁶ In the Americas cohort, spironolactone produced an HR of 0.82 (0.69–0.98). A subsequent metabolite substudy found that approximately 30% of patients enrolled from Russia had no detectable canrenone - the active metabolite of spironolactone - in their plasma samples.⁷ In a substantial proportion of those patients, the drug had apparently not been taken, or not been dispensed at all.

TOPCAT is not primarily a lesson about endpoint design. But it sits alongside these methodological concerns as a reminder that an endpoint can be perfectly defined on paper and still be rendered uninterpretable by the patients enrolled beneath it.

The Third Problem: We Were Not Asking the Right Question

Behind the undercounting of recurrent events and the invisibility of outpatient deterioration, there is a question the conventional composite endpoint was never really built to answer: what actually matters to this patient?

The traditional composite treats death, hospitalization, and everything in between as a single pooled outcome. A patient who dies three days after enrollment contributes the same statistical weight as a patient hospitalized five years later. A trial can read out positive because it reduced hospitalizations while leaving quality of life unchanged. Or it can miss a genuine benefit entirely because the primary signal was symptomatic rather than event-based. Neither scenario is hypothetical.

Win ratio, introduced formally by Pocock and colleagues in 2012, tries to restructure the question.¹⁰ Instead of pooling all events into a rate, it compares each treated patient against each control patient in the following order: who died first? If neither, who had more hospitalizations? If still tied, who felt better by symptom score? The result - a ratio of “wins” to “losses” for the treatment arm - preserves clinical hierarchy in a way that traditional composites flatten.

EMPULSE applied this in acute heart failure for the first time. In 530 patients, empagliflozin produced a win ratio of 1.36 (95% CI 1.09–1.68) across mortality, heart failure events, and KCCQ score at 90 days.⁸ Small trial, short follow-up - this is proof-of-concept. But it showed that a single endpoint could integrate hard events and patient-reported symptom burden, weighted in the order that a clinician would actually prioritize them.

STEP-HFpEF took the method into territory where a conventional endpoint would have been genuinely blind. In 529 patients with obesity-related HFpEF, semaglutide produced a win ratio of 1.72 (95% CI 1.37–2.15).⁹ Only 1 heart failure hospitalization occurred in the semaglutide arm versus 12 in the placebo arm - far too few events for any hospitalization-based endpoint to have power. The benefit in this population was predominantly symptomatic. A traditional composite would have returned a null result and reported nothing. Win ratio made the actual clinical effect visible.

The method has real critics, and the criticism is substantive. Butler, Stockbridge, and Packer raised concerns in Circulation about win ratio’s sensitivity to censoring distributions, the somewhat arbitrary tie-breaking threshold for symptom scores, and the difficulty of translating a win ratio into an absolute risk estimate that clinicians can act on.¹¹ These are not trivial objections. Whether win ratio becomes a standard endpoint in heart failure trials or remains a useful tool in specific phenotypes is not a settled question.

What pushed the field to explore it anyway was the recognition that the endpoints we had were leaving things out - not occasionally, but systematically, in ways that tracked with how the disease behaved and how care had evolved.

Subscribe now

Three Questions for Reading a Heart Failure Trial

The practical implication is not that trial results should be distrusted. It is that they deserve a specific kind of attention to what was actually being measured.

When a heart failure trial reports its primary outcome, three questions are worth holding onto.

Was this a time-to-first-event analysis? If so, the published HR is probably a conservative estimate - particularly in populations prone to multiple hospitalizations. The drug may have been doing more than the headline number captured.

Did the endpoint include urgent visits requiring intravenous therapy? If not, the trial was partly measuring a world that clinical practice has moved on from. In contemporary outpatient management, the real burden of worsening heart failure extends well beyond formal admissions.

Were the patients in the trial representative of the patient in front of you? TOPCAT is the permanent reminder that enrollment quality is a form of internal validity. An endpoint that is correctly defined can still be rendered uninterpretable by who ends up underneath it.

These are not exotic statistical concerns. They are reading skills - the kind that change how a published number becomes a clinical decision.

The endpoints are getting better because they had to. Heart failure forced the question.

—Cardio Concepts

References

1. Mogensen UM, et al. Recurrent hospitalizations in patients with heart failure: reanalysis of PARADIGM-HF. Eur J Heart Fail. 2018. https://doi.org/10.1002/ejhf.1139

2. Rogers JK, et al. Analysing recurrent hospitalizations in heart failure: a review of statistical methodology, with application to CHARM-Preserved. Eur J Heart Fail. 2014. https://doi.org/10.1002/ejhf.29

3. Okumura N, et al. Worsening heart failure: the clinical phenotype of a deteriorating patient with chronic heart failure. Circulation. 2016. https://doi.org/10.1161/CIRCULATIONAHA.115.020729

4. Bhatt DL, et al. (SOLOIST-WHF). Sotagliflozin in patients with diabetes and recent worsening heart failure. N Engl J Med. 2021. https://doi.org/10.1056/NEJMoa2030183

5. Solomon SD, et al. (DELIVER). Dapagliflozin in heart failure with mildly reduced or preserved ejection fraction. N Engl J Med. 2022. https://doi.org/10.1056/NEJMoa2206286

6. Pfeffer MA, et al. (TOPCAT regional). Regional variation in patients and outcomes in the TOPCAT trial. Circulation. 2015. https://doi.org/10.1161/CIRCULATIONAHA.114.013255

7. de Denus S, et al. (TOPCAT canrenone substudy). Spironolactone metabolites in TOPCAT — new insights into regional variation. N Engl J Med. 2017. https://doi.org/10.1056/NEJMc1612601

8. Voors AA, et al. (EMPULSE). Empagliflozin in patients hospitalized for acute heart failure: a multinational randomized trial. Nat Med. 2022. https://doi.org/10.1038/s41591-021-01659-1

9. Kosiborod MN, et al. (STEP-HFpEF). Semaglutide in patients with heart failure with preserved ejection fraction and obesity. N Engl J Med. 2023. https://doi.org/10.1056/NEJMoa2306963

10. Pocock SJ, et al. The win ratio: a new approach to the analysis of composite endpoints in clinical trials based on clinical priorities. Eur Heart J. 2012. https://doi.org/10.1093/eurheartj/ehr352

11. Butler J, Stockbridge N, Packer M. Reassessing the clinical trial endpoint framework for heart failure. Circulation. 2024. https://doi.org/10.1161/CIRCULATIONAHA.123.067786

    Share

    Leave a comment

    Explore the Cardio Concepts Library

There is something quietly uncomfortable about the way cardiology has treated its oldest drugs. We reach for them when we run out of better ideas, or we abandon them the moment a trial comes up negative, as if a single study settles what decades of clinical experience opened up. The five drugs, or drug classes, at the center of this essay don’t fit neatly into either category. They are too old to be exciting, too pharmacologically interesting to be dismissed, and too entangled with our own evolving understanding of heart failure and coronary disease to be summarized cleanly.

What follows is not a review of what each drug does. Most clinicians already know the mechanism, the trials, the guideline footnotes. What’s worth exploring is something harder to articulate: why these drugs keep coming back, and what that return reveals about the way we ask questions in cardiovascular medicine.

Digitalis and the Problem of the Wrong Endpoint

The story of cardiac glycosides in heart failure is often told as a cautionary tale. DIG (1997) showed digoxin didn’t reduce mortality. Clinical inertia carried it forward for years, then guidelines quietly pushed it to the margins. End of story.

Except it wasn’t.

DIG showed something more specific than “digoxin doesn’t work.” It showed digoxin - at doses and serum concentrations frequently above 1.0 ng/mL, in an era before quadruple therapy existed - didn’t reduce all-cause mortality. It did reduce hospitalization. That signal got buried under the mortality headline, and with it, a more careful question: does the glycoside mechanism have residual value in a patient already on a beta-blocker, an ACE inhibitor, an MRA, and an SGLT2 inhibitor? That question wasn’t askable in 1997 because that clinical context didn’t exist.

DIGIT-HF asked it. Digitoxin, not digoxin, added on top of modern guideline-directed therapy in HFrEF patients receiving contemporary background treatment. The choice of digitoxin was deliberate: hepatic metabolism independent of renal clearance, longer half-life, and a narrower toxicity window at therapeutic concentrations made it a cleaner test of the glycoside hypothesis. The result was an 18% reduction in the composite of cardiovascular death and heart failure hospitalization (HR 0.82; 95% CI 0.69–0.98) - the first positive cardiac glycoside trial in 28 years.

The post-hoc analysis by AF status (Bavendiek et al., Eur Heart J 2026) added clinical texture. In HFrEF patients with concomitant atrial fibrillation, digitoxin’s dual mechanism - modest positive inotropy alongside vagally-mediated rate control - may carry particular weight. Beta-blockers in HFrEF with AF have consistently shown less mortality benefit than in sinus rhythm; this isn’t new, but it has rarely been used to think about what to reach for next. If digitoxin offers rate control without the negative chronotropic penalty in a dysfunctional ventricle, that distinction has practical consequences.

DIG’s shadow lasted nearly three decades partly because we treated a primary endpoint as a final verdict on a mechanism. The biology of glycoside inhibition didn’t change. The clinical context being tested did - and that turned out to matter enormously.

H-ISDN and the Race Phenotype Problem

The hydralazine-isosorbide dinitrate story is simultaneously one of the most important and most misread chapters in heart failure pharmacology.

V-HeFT I (1986) was the first randomized trial to show mortality reduction with a vasodilator in heart failure - predating ACE inhibitor dominance. V-HeFT II (1991) showed enalapril was superior, and that result appropriately changed practice. But the comparison wasn’t symmetrical. Enalapril winning didn’t mean H-ISDN had no effect; it meant enalapril had more. In patients intolerant of ACE inhibitors or ARBs, the combination retained biological plausibility.

A post-hoc analysis of V-HeFT then identified differential treatment response by self-identified race, with Black patients appearing to derive greater benefit from H-ISDN relative to enalapril than white patients. This drove A-HeFT (2004), which enrolled exclusively self-identified Black patients with HFrEF and showed a 43% reduction in mortality - a result striking enough that the trial was stopped early. BiDil received FDA approval in 2005, the first drug approved for a race-defined indication.

The approval has been contentious, and not without reason. Self-reported race is a social construct that correlates imperfectly with genetic ancestry, and ancestry correlates imperfectly with the molecular phenotypes that determine drug response. The risk is that a demographic label ends up doing the work that a biological measurement should be doing.

The biology underlying the racial signal in A-HeFT points toward nitric oxide bioavailability, oxidative stress burden, and NADPH oxidase activity - mechanisms that differ in prevalence across populations but are not defined by census categories. Hydralazine’s capacity to inhibit NADPH oxidase and reduce superoxide-mediated NO degradation is the pharmacologically active hypothesis. If correct, the relevant patient for H-ISDN is someone with high oxidative stress, low NO bioavailability, and a neurohormonal milieu where conventional RAS blockade is insufficient or not tolerated - a phenotype that exists across demographic groups, and probably goes undertreated because it doesn’t match a racial label on a prescription form.

A-HeFT’s finding was real in the population studied. But using race as the selection criterion was always a pragmatic shortcut, not a biological answer. That more precise question - who actually has the oxidative stress phenotype that makes H-ISDN mechanistically relevant - still hasn’t been answered in a way that changes clinical practice.

Ivabradine and the Meaning of Heart Rate

Heart rate carries two interpretations that matter clinically. It is a marker - of autonomic tone, sympathetic activation, disease severity. And it is a mediator - in the failing or ischemic heart, each additional beat carries an oxygen cost, increases wall stress, shortens diastolic filling time. Whether HR is primarily a marker or a mediator in a given patient determines whether lowering it helps.

Ivabradine made this distinction pharmacologically testable. The specificity was supposed to be an advantage: lower heart rate without touching contractility, blood pressure, or AV conduction. In dilated ventricles with volume overload, it was an advantage.

SHIFT (2010) was the clearest demonstration: HFrEF with resting HR ≥70 bpm on background therapy including beta-blockers. Ivabradine reduced the composite of cardiovascular death and heart failure hospitalization by 18%. The mechanism was coherent - rate reduction extends diastolic filling time, reduces myocardial oxygen demand, and over time may contribute to reverse remodeling.

SIGNIFY (2014) enrolled a different patient: stable coronary disease without LV dysfunction, higher HR threshold, higher dose. Not only was the primary outcome neutral - in the pre-specified angina subgroup there was a signal toward harm. In a ventricle with preserved systolic function and chronotropic reserve, slowing rate reduces cardiac output enough to trigger reflex sympathetic counter-regulation. That counter-regulation appears to offset, and in some patients potentially reverse, whatever benefit rate reduction might have provided.

EDIFY (2017) applied the same logic to HFpEF: a stiff, non-dilated ventricle where rate is partly compensating for impaired stroke volume. No meaningful benefit, consistent with the same underlying principle.

The clinical implication is not complicated to state but harder to operationalize: ivabradine is useful when tachycardia is driving harm, not when it’s compensating for something else. Telling the difference requires thinking about ventricular geometry and filling dynamics, not just reading a number off the chart.

Spironolactone and the Trial That Lied

RALES (1999) shifted how the field thought about aldosterone. Before it, aldosterone was mostly managed as a side effect of renin-angiotensin activation - a renal hormone relevant to potassium and sodium balance. After RALES, aldosterone became a driver of myocardial fibrosis, endothelial dysfunction, and sympathetic activation in its own right. Spironolactone at low dose in severe HFrEF reduced all-cause mortality by 30%, and the mechanism wasn’t primarily about electrolytes.

Extending this to HFpEF made biological sense. HFpEF carries significant myocardial fibrosis as a substrate. Aldosterone levels are elevated. MRA blockade should, in principle, modify the fibrotic trajectory. TOPCAT (2014) was designed to test this. It came up neutral.

And then it turned out that TOPCAT was, in a meaningful sense, two different trials running simultaneously.

The analysis by de Denus et al. (NEJM 2017) was methodologically damning. Patients enrolled in Russia and Georgia - roughly a third of the TOPCAT population - had no detectable spironolactone metabolites in stored serum samples. Not lower levels: no detectable metabolite. These patients were randomized to spironolactone and apparently received either nothing or a pharmacologically inactive product. When the analysis was restricted to patients from the Americas, where adherence was biochemically confirmed, the hazard ratio moved into positive territory with confidence intervals no longer crossing one.

It’s a fairly rare occurrence: a trial’s primary result being largely attributable to nonadherence in a geographically defined subgroup, confirmed post-hoc by serum biomarker analysis. TOPCAT’s neutral headline spent a decade shaping practice in HFpEF, and it was doing so on the basis of a structurally compromised dataset.

FINEARTS-HF (2024) then asked whether finerenone - a non-steroidal, selective MRA with a different safety and tissue-selectivity profile - could confirm the MRA hypothesis in HFpEF under conditions of verified adherence and contemporary trial design. The composite of cardiovascular death and worsening heart failure events was reduced (HR 0.84; 95% CI 0.74–0.95), in a population defined by objective congestion criteria and elevated natriuretic peptides.

The underlying hypothesis was never wrong. It just needed a trial that actually tested it.

Colchicine and the Timing Problem

CANTOS (2017) proved something the field had long suspected but struggled to demonstrate cleanly: residual cardiovascular risk in patients on statin therapy is partly driven by inflammation, and targeting that inflammation with a specific agent - canakinumab, an IL-1β antibody - reduces events independent of LDL. The mechanism was convincing. The drug was not going to be widely used at its price point.

COLCOT (2019) and LoDoCo2 (2020) showed that colchicine at 0.5 mg daily could reduce major adverse cardiovascular events in post-MI patients and in stable chronic coronary syndrome, respectively. The biological rationale was convergent with CANTOS - colchicine dampens NLRP3 inflammasome activity and neutrophil-mediated vascular inflammation, the same low-grade inflammatory axis responsible for residual risk in patients already on optimal lipid-lowering therapy. The results were consistent, the dose was low, the drug was cheap.

The remaining question was acute-phase intervention. If chronic coronary inflammation drives residual risk, could earlier colchicine - started within 72 hours of STEMI - do even more? CLEAR-SYNERGY (2025) tested exactly this. Completely neutral.

The explanation probably comes down to what the inflammatory environment actually looks like in the first 72 hours of a STEMI. This is not simply elevated IL-1β signaling on top of a stable baseline. It is a response driven by cell necrosis, complement activation, massive neutrophil infiltration, and multiple redundant pro-inflammatory cascades running in parallel. Colchicine’s mechanism is well-matched to the chronic, lower-intensity inflammatory milieu of stable coronary artery disease, where any single pathway is more likely to be rate-limiting. In the acute setting, inhibiting one node doesn’t move the system enough to show up in clinical outcomes.

What CLEAR-SYNERGY clarifies is that “anti-inflammatory” is not a property that transfers uniformly across disease phases. The inflammatory state in stable CCS and the inflammatory state 48 hours post-STEMI are biologically different enough that a drug effective in one may have no leverage in the other. This distinction was somewhat predictable from first principles, but it took a 7,000-patient trial to establish it empirically.

Subscribe now

What These Five Stories Share

Each of these drugs was, at some point, declared to have failed. Digoxin failed DIG. H-ISDN lost to enalapril. Ivabradine failed SIGNIFY and EDIFY. Spironolactone failed TOPCAT. Colchicine failed CLEAR-SYNERGY.

In each case, the failure turned out to be more specific than the headline suggested. DIG asked the wrong question in the wrong era. TOPCAT was structurally compromised. SIGNIFY and EDIFY enrolled the wrong phenotype. CLEAR-SYNERGY intervened at the wrong biological moment. H-ISDN’s racial label was a proxy for a biological question not yet asked precisely enough.

The pattern reveals something about how we read trials. We are good at using primary outcomes to close questions. We are less good at using them to narrow questions. A negative primary endpoint tends to become categorical - “this drug doesn’t work” - when it should often mean something more specific: this drug, at this dose, in this population, at this stage of disease, didn’t work.

That difference has had real consequences. Patients with HFrEF and AF went without a glycoside option for nearly three decades. HFpEF patients were arguably denied a meaningful therapeutic hypothesis for a decade based on a contaminated dataset. Patients with stable coronary disease and measurable inflammatory burden may still be undertreated with colchicine because CLEAR-SYNERGY’s result is being read without the timing discrimination it actually supports.

The task is not rehabilitating old drugs for its own sake. It is learning to ask better biological questions - about which mechanism is rate-limiting in which patient, at which point in the disease course - and resisting the reflex to treat a trial’s primary outcome as the last word on a pharmacological hypothesis. Some drugs that failed broad trials will prove useful in narrower, better-defined populations. Others genuinely won’t. The work is building the clinical reasoning to tell the difference.

If this perspective resonates with you, I’ve made the material available in two formats:

• A free PDF for reading and sharing.

• A fully editable PowerPoint deck with complete speaker notes and references for those who would like to explore the evidence in greater depth or use it for teaching.

I hope it proves useful in your own clinical thinking, discussions, or lectures.

—Cardio Concepts

References

1. Digitalis Investigation Group. The effect of digoxin on mortality and morbidity in patients with heart failure. N Engl J Med. 1997;336:525–533. DOI: 10.1056/NEJM199702203360801

2. Bavendiek U, et al. Digitoxin in patients with heart failure and reduced ejection fraction (DIGIT-HF): a randomised, double-blind, placebo-controlled trial. N Engl J Med. 2025. DOI: 10.1056/NEJMoa2415471

3. Bavendiek U, et al. Effect of digitoxin by atrial fibrillation status in heart failure with reduced ejection fraction: post-hoc analysis of DIGIT-HF. Eur Heart J. 2026. DOI: 10.1093/eurheartj/ehag379

4. Cohn JN, et al. Effect of vasodilator therapy on mortality in chronic congestive heart failure (V-HeFT I). N Engl J Med. 1986;314:1547–1552. DOI: 10.1056/NEJM198606123142404

5. Cohn JN, et al. A comparison of enalapril with hydralazine–isosorbide dinitrate in the treatment of chronic congestive heart failure (V-HeFT II). N Engl J Med. 1991;325:303–310. DOI: 10.1056/NEJM199108013250502

6. Taylor AL, et al. Combination of isosorbide dinitrate and hydralazine in blacks with heart failure (A-HeFT). N Engl J Med. 2004;351:2049–2057. DOI: 10.1056/NEJMoa042934

7. Fox K, et al. Ivabradine for patients with stable coronary artery disease and left-ventricular systolic dysfunction (BEAUTIFUL). Lancet. 2008;372:807–816. DOI: 10.1016/S0140-6736(08)61170-8

8. Swedberg K, et al. Ivabradine and outcomes in chronic heart failure (SHIFT). Lancet. 2010;376:875–885. DOI: 10.1016/S0140-6736(10)61198-1

9. Fox K, et al. Ivabradine in stable coronary artery disease without clinical heart failure (SIGNIFY). N Engl J Med. 2014;371:1091–1099. DOI: 10.1056/NEJMoa1406430

10. Komajda M, et al. Effect of ivabradine in patients with heart failure and preserved ejection fraction: the EDIFY randomized placebo-controlled trial. Eur J Heart Fail. 2017;19:1495–1503. DOI: 10.1002/ejhf.876

11. Pitt B, et al. The effect of spironolactone on morbidity and mortality in patients with severe heart failure (RALES). N Engl J Med. 1999;341:709–717. DOI: 10.1056/NEJM199909023411001

12. Pitt B, et al. Spironolactone for heart failure with preserved ejection fraction (TOPCAT). N Engl J Med. 2014;370:1383–1392. DOI: 10.1056/NEJMoa1313731

13. de Denus S, et al. Spironolactone metabolites in TOPCAT — new insights into regional variation. N Engl J Med. 2017;376:1690–1692. DOI: 10.1056/NEJMc1612601

14. Solomon SD, et al. Finerenone in heart failure with mildly reduced or preserved ejection fraction (FINEARTS-HF). N Engl J Med. 2024;391:1475–1485. DOI: 10.1056/NEJMoa2407107

15. Ridker PM, et al. Antiinflammatory therapy with canakinumab for atherosclerotic disease (CANTOS). N Engl J Med. 2017;377:1119–1131. DOI: 10.1056/NEJMoa1707914

16. Tardif JC, et al. Efficacy and safety of low-dose colchicine after myocardial infarction (COLCOT). N Engl J Med. 2019;381:2497–2505. DOI: 10.1056/NEJMoa1912388

17. Nidorf SM, et al. Colchicine in patients with chronic coronary disease (LoDoCo2). N Engl J Med. 2020;383:1838–1847. DOI: 10.1056/NEJMoa2021372

18. Jolly SS, et al. Colchicine in acute myocardial infarction (CLEAR-SYNERGY). N Engl J Med. 2025;392:633–642. DOI: 10.1056/NEJMoa2405922

    Share

    Leave a comment

There is something quietly remarkable about colchicine. A drug extracted from autumn crocus, used for gout since the Byzantine era, now sitting in cardiology guidelines with an FDA approval specifically for cardiovascular risk reduction. It got there not through drug discovery, but because someone asked a different question about where cardiovascular events actually come from after the lipids have been treated.

That question - whether the inflammatory component of atherosclerosis is independently targetable and clinically meaningful - is what this essay is really about. Colchicine is the vehicle. The bigger story is what the trials collectively tell us about residual cardiovascular risk and how specific we need to be about who we are actually treating.

Setting the Stage: CANTOS and the Proof of Concept

To understand colchicine’s trajectory in cardiology, you have to start with canakinumab. CANTOS (2017) was the trial that changed the conceptual framework - not because canakinumab became a treatment, but because it proved the underlying biology.

The premise was specific: if you take post-MI patients with persistent elevated hsCRP (≥2 mg/L) - people already on statins, already treated - can you reduce cardiovascular events by targeting IL-1β directly? The answer was yes. Canakinumab 150 mg reduced the composite of CV death, MI, and stroke with an HR of 0.85. LDL didn’t move. The mechanism was purely anti-inflammatory.

That finding reoriented thinking. Inflammation wasn’t just a biomarker or a parallel phenomenon in atherosclerosis. It was a targetable pathway with measurable clinical consequences. The problem was the drug itself: around $200,000 per year, and a signal of increased fatal infection that was real enough to be concerning. CANTOS proved the concept and buried the drug in the same breath.

But it opened a door. Colchicine works partly through NLRP3 inflammasome inhibition - the same pathway that sits upstream of IL-1β production - via tubulin polymerization inhibition and neutrophil trafficking disruption. Cheap, oral, decades of safety data in gout and pericarditis. The mechanistic overlap with the CANTOS pathway was compelling enough that the next logical question was whether it translated into clinical outcomes.

COLCOT: The First Signal

COLCOT (2019) brought colchicine into cardiology conversations in a serious way. Post-MI patients, within 30 days of the event, randomized to colchicine 0.5 mg daily versus placebo. Median follow-up around 22 months. The primary composite - CV death, cardiac arrest, MI, stroke, urgent revascularization - came in at HR 0.77. An absolute risk reduction of 1.6%, NNT around 63 over two years.

On the surface, a clean positive trial. But the details deserve more than a headline.

The fragility index was 9. Nine events in the other direction and the result flips to nonsignificant. The benefit was driven substantially by urgent revascularization and stroke - the softer endpoints in the composite. The stroke result was striking on its own: HR 0.26, a 74% relative reduction. But stroke as a primary driver in a post-MI trial is an unusual finding, and when a composite moves mainly on its softer components, the robustness of the overall signal is a fair question.

There was also a pneumonia signal - 0.9% versus 0.4% - consistent with colchicine’s immunomodulatory effects. Not a reason to dismiss the trial, but something to carry forward.

COLCOT was important as a first RCT. What it couldn’t do, with those numbers and that fragility index, was close the question.

LoDoCo2: A Different Population, a More Convincing Answer

LoDoCo2 (2020) didn’t just test colchicine in a larger group. It tested it in a different clinical context, and that distinction turned out to matter.

The population: stable chronic coronary syndrome, at least six months from any acute event. Not the inflammatory surge of acute MI, but the smoldering, persistent, low-grade inflammation that characterizes established atherosclerosis in someone who is otherwise clinically stable. Colchicine 0.5 mg daily versus placebo, median follow-up 28.6 months, with a run-in phase to exclude non-tolerators from the start.

The result: HR 0.69, NNT approximately 36 over 2.5 years. This became the basis for the FDA approval of Lodoco in June 2023 - the first anti-inflammatory drug approved specifically for cardiovascular risk reduction in chronic coronary syndrome.

The caveats are real. The run-in enriches the population for tolerators, so real-world NNTs will be higher. There was a non-significant trend toward higher non-CV mortality in the colchicine arm (HR 1.51, P=0.059) - unexplained, not replicated elsewhere, but present enough to notice. Long-term data from Opstal et al. in Circulation (2022) showed the benefit persisting to year four (HR 0.53 for the primary endpoint), which is reassuring - if a non-CV mortality signal were accumulating, you’d expect it to appear there.

What LoDoCo2 defined was a phenotype: stable CAD, chronic inflammation, not the acute phase. That specificity is part of why this remains the strongest dataset for colchicine in cardiovascular disease.

Where Colchicine Doesn’t Work: The Stroke and Acute-Phase Data

CONVINCE (2024) tested colchicine after non-cardioembolic stroke or TIA. Reasonable hypothesis, problematic execution - stopped early, significantly underpowered, open-label PROBE design. The ITT result was HR 0.84 with a confidence interval crossing 1.0. Neutral.

CHANCE-3 (2024), in Chinese patients with acute ischemic stroke or TIA: HR 0.98. No signal at all.

Neither result is particularly surprising once you think through the mechanism. Colchicine’s apparent benefit in LoDoCo2 comes from suppressing sustained, low-grade vascular inflammation in established atherosclerosis. Acute cerebrovascular events involve different pathophysiology - atherothrombosis, small vessel disease, cardioembolic mechanisms, acute inflammatory cascades that differ from the chronic plaque biology colchicine seems to address. The drug may not be targeting the right biology in that context.

What the stroke data clarifies is that colchicine isn’t a broad cardiovascular anti-inflammatory. It appears to reduce events in a specific context, through a specific mechanism, in a specific population - and stroke, at least at current evidence, isn’t that population.

CLEAR-SYNERGY: The Result Nobody Fully Expected

In early 2025, CLEAR-SYNERGY published its colchicine arm results.

The trial enrolled patients undergoing PCI after acute MI - predominantly STEMI, within 72 hours. N=7,062. Follow-up approximately three to four years. The primary composite of CV death, MI, stroke, and ischemia-driven revascularization: HR 0.99. P=0.93. Completely neutral.

CRP levels fell in the colchicine arm. The anti-inflammatory effect was biochemically real. It just didn’t translate into outcomes.

The contrast with COLCOT is hard to set aside. Same drug, same indication on paper, opposite results. The natural question is whether these trials are actually testing the same thing.

Some differences are meaningful. CLEAR-SYNERGY enrolled within 72 hours of MI - the peak of acute inflammatory surge. COLCOT enrolled up to 30 days post-event, after the acute phase had partly resolved. CLEAR-SYNERGY was predominantly STEMI; COLCOT was more mixed. And CLEAR-SYNERGY was larger and longer, which generally makes it more reliable, not less.

One hypothesis that holds up mechanistically: timing matters more than we appreciated. The acute post-STEMI inflammatory response is massive - IL-1, IL-6, and CRP spike dramatically in the first 72 hours, driven by a cascade that may be biologically different from the chronic, smoldering NLRP3-mediated inflammasome activation in stable atherosclerotic plaque. Colchicine may be suppressing one and not the other.

A less comfortable reading: COLCOT was a fragility-index-9 trial that crossed the significance threshold, and CLEAR-SYNERGY - larger, longer, more carefully powered - is closer to the truth about colchicine in the post-MI setting.

The honest answer is that both interpretations are consistent with the current data, and the field hasn’t resolved this. Any practice pattern that ignores that uncertainty is moving faster than the evidence.

What This Means in Practice

The practical question - who actually gets colchicine - has gotten harder to answer after CLEAR-SYNERGY, not easier.

For stable chronic coronary syndrome, the case is still reasonably intact. LoDoCo2 is solid. The FDA approval reflects real data. For a patient with established CAD, on optimal medical therapy, with evidence of persistent inflammatory activity - whether that’s an elevated hsCRP, a history of recurrent events despite treatment, or a clinical phenotype suggesting residual risk - colchicine 0.5 mg daily is defensible. The drug is inexpensive and the NNT is competitive with other secondary prevention interventions. Practically, GI tolerability is the main barrier; taking it with food helps considerably. The infectious risk is real but modest, and it matters more in patients who are already immunocompromised or on interacting drugs.

For post-MI, the evidence is now genuinely unsettled. The 2023 ESC guidelines included colchicine at a IIb recommendation level, written before CLEAR-SYNERGY published. Whether and how that recommendation evolves is an open question. My read is that post-MI colchicine, if used, should be reserved for patients with identifiable high inflammatory burden - persistently elevated hsCRP despite statin therapy, or a phenotype with recurrent events - rather than as a reflex add-on after any MI.

The phenotyping problem underneath all of this is unsolved. We don’t have reliable tools to identify which patients carry clinically significant residual inflammatory risk. hsCRP is accessible but imperfect. The optimal threshold, measurement timing, and how to translate biomarker values into individual treatment decisions remain empirical rather than evidence-based in any precise sense.

Subscribe now

The Larger Point

Lipid-lowering is mature. High-intensity statins, ezetimibe, PCSK9 inhibitors - that pathway is well-worked. Patients still have events. Atherosclerosis continues. CANTOS established that inflammation is a parallel, independently targetable driver - and colchicine is the pragmatic attempt to access that biology at a price and with a delivery mechanism that actually makes clinical sense.

What the colchicine trials collectively show is that “does this drug work?” is too coarse a question. The better question is: in which patient, at which point in the disease course, targeting which aspect of vascular inflammation? LoDoCo2 gives a partial answer. CLEAR-SYNERGY narrows it. The stroke trials close off one direction. The picture that emerges isn’t a failure - it’s a more honest map of where the drug’s mechanism is and isn’t relevant.

Colchicine didn’t change. What changed is our understanding of when the biology it modulates is the biology that’s actually driving events in a given patient. Getting that right - phenotyping before prescribing - is probably where the real work in cardiovascular inflammation sits for the next decade.

I've put together this one-page framework summarizing where colchicine fits in coronary artery disease, from acute MI to stable CAD, through the lens of inflammatory biology rather than diagnosis. Feel free to download it for free if you find it useful.

Colchicine in CAD

—Cardio Concepts

Previous The TOPCAT Problem: What a Flawed Trial Taught Us About HFpEF, Aldosterone, and the Limits of Evidence

Next Old Drugs, New Questions: Five Stories About What We Got Wrong the First Time

References

1. Ridker PM, Everett BM, Thuren T, et al.; CANTOS Trial Group. Antiinflammatory therapy with canakinumab for atherosclerotic disease. N Engl J Med. 2017;377(12):1119–1131. doi:10.1056/NEJMoa1707914

2. Tardif JC, Kouz S, Waters DD, et al. Efficacy and safety of low-dose colchicine after myocardial infarction. N Engl J Med. 2019;381(26):2497–2505. doi:10.1056/NEJMoa1912388

3. Nidorf SM, Fiolet ATL, Mosterd A, et al.; LoDoCo2 Trial Investigators. Colchicine in patients with chronic coronary disease. N Engl J Med. 2020;383(19):1838–1847. doi:10.1056/NEJMoa2021372

4. Opstal TSJ, Fiolet ATL, van Broekhoven A, et al. Colchicine in patients with chronic coronary disease — long-term follow-up. Circulation. 2022;145:626–633. doi:10.1161/CIRCULATIONAHA.121.057595

5. Kelly P, Lemmens R, Weimar C, et al. Colchicine for prevention of vascular events in ischaemic stroke and transient ischaemic attack (CONVINCE). Lancet. 2024;404(10448):125–133. doi:10.1016/S0140-6736(24)00968-1

6. Li J, Meng X, Shi FD, et al. Colchicine for secondary prevention after ischaemic stroke or transient ischaemic attack (CHANCE-3). BMJ. 2024;385:e079061. doi:10.1136/bmj-2023-079061

7. Jolly SS, d’Entremont M-A, Lee SF, et al.; CLEAR Investigators. Colchicine in acute myocardial infarction (CLEAR-SYNERGY/OASIS-9). N Engl J Med. 2025;392:633–642. doi:10.1056/NEJMoa2405922

   Share

   Leave a comment

   Explore the Cardio Concepts Library

There’s a particular kind of frustration that comes with reading a negative trial when you believe, based on mechanism, on physiology, on everything you understand about the disease, that the drug should have worked.

TOPCAT is that trial for most of us who care about HFpEF.

Published in 2014, TOPCAT randomized 3,445 patients with HFpEF to spironolactone or placebo. The primary composite - cardiovascular death, cardiac arrest, or hospitalization for heart failure - came back neutral: HR 0.89 (95% CI 0.77–1.04), P=0.14. And with that, a lot of people concluded that aldosterone blockade didn’t work in HFpEF, and moved on.

Except that’s not really what happened. And the story of why it’s not, and what we’ve learned from picking that trial apart over the past decade, tells us something important not just about spironolactone, but about how we read evidence in a disease as heterogeneous as HFpEF.

Why Aldosterone Was a Reasonable Target in the First Place

To understand why TOPCAT mattered, you have to go back to RALES.

In 1999, RALES was a different kind of trial - straightforward in its logic and dramatic in its results. Patients with severe HFrEF (LVEF ≤35%, NYHA III–IV), already on ACE inhibitors and diuretics, were randomized to spironolactone 25 mg daily. The trial was stopped early after a median of 24 months because the mortality benefit was too large to ignore: 35% versus 46% all-cause mortality, RR 0.70, NNT of approximately 9. Hospitalization for heart failure dropped by 35%.

RALES established spironolactone as the third pillar of GDMT in HFrEF, alongside ACE inhibitors and beta-blockers. But more importantly for what came after, it also reframed how we thought about aldosterone - not just as a diuretic target, but as a mediator of fibrosis, inflammation, and myocardial remodeling. Collagen deposition, ventricular stiffening, neurohormonal activation: these effects are relevant well beyond the fluid-overloaded HFrEF patient.

That reasoning is why HFpEF looked like a logical next frontier. The core problem in HFpEF is diastolic dysfunction - impaired relaxation and increased chamber stiffness, driven substantially by interstitial fibrosis and extracellular matrix accumulation. Aldosterone promotes exactly that. The biology connected.

ALDO-DHF, published in JAMA in 2013, gave the mechanism some clinical traction. In 422 patients with HFpEF (LVEF ≥50%, NYHA II/III), twelve months of spironolactone 25 mg significantly improved E/e’ by 1.5 units compared to a slight increase in the placebo group (P<0.001). LV mass index also decreased. The echocardiographic signal was real.

What ALDO-DHF didn’t show was any improvement in peak VO₂ or six-minute walk distance. Exercise capacity stayed flat. That gap - structural improvement without functional benefit - is one of the recurring frustrations of HFpEF research, and it should have tempered expectations somewhat going into TOPCAT. But the question of whether aldosterone blockade could reduce hard outcomes still seemed worth asking.

What TOPCAT Actually Found - and Didn’t Find

The neutral primary endpoint got most of the attention, but a few things in the main paper are worth pausing on. Hospitalization for heart failure - a pre-specified secondary endpoint - was significantly reduced: HR 0.83 (95% CI 0.69–0.99), P=0.04. In a trial of this size and duration, that’s a clinically meaningful signal, not a rounding error.

But the more important finding came from the regional data. TOPCAT enrolled at 270 sites across six countries, including a substantial cohort from Russia and Georgia. When Pfeffer and colleagues published the regional variation analysis in Circulation in 2015, the numbers were striking: in the Americas sites, the placebo event rate over four years was 31.8%. In Russia and Georgia, it was 8.4%. A fourfold difference, within the same trial, with the same enrollment criteria.

That kind of heterogeneity doesn’t have a clean explanation. It means either the patients were fundamentally different, different phenotypes, different disease severity, or the data from some sites had serious problems. When you looked at the Americas subgroup in isolation, spironolactone was effective: HR 0.82 (95% CI 0.69–0.98), P=0.026. The drug appeared to work in the patients who were actually sick enough to have events.

What followed was the de Denus metabolite analysis in 2017. The investigators measured canrenone, the primary active metabolite of spironolactone, in stored serum samples from TOPCAT participants. Among patients from Russia and Georgia who had been randomized to the spironolactone arm, roughly 30% had no detectable canrenone. Not low levels. None. Meaning a substantial proportion of patients assigned to active treatment almost certainly never took the drug.

At that point the neutral primary endpoint stops being a statement about spironolactone’s efficacy. It becomes a statement about what happens when you average a real treatment effect against a contaminated control condition - which is that you get noise.

The Harder Question Underneath

The integrity issue explains the neutral result. But it also invites a more uncomfortable question: even if TOPCAT had been perfectly conducted, would spironolactone have worked across every patient enrolled?

Probably not - and that’s worth sitting with.

HFpEF is not a single disease. It’s a clinical label that collects a range of phenotypes differing substantially in their underlying pathophysiology. Aldosterone-mediated fibrosis is one driver among many. In a patient whose HFpEF is primarily driven by obesity-related systemic inflammation, or by atrial fibrillation and volume-mediated diastolic impairment, or by infiltrative disease - blocking aldosterone may address a secondary contributor at best. Expecting consistent benefit across that entire population was optimistic, regardless of what happened in Russia and Georgia.

This is the structural problem with large, broadly enrolled HFpEF trials. You’re averaging across phenotypes. When a therapy is highly relevant for some biological contexts and largely irrelevant for others, you get diluted results - neutral primaries with signals buried in subgroups - even when the drug genuinely works in the right patient. The Americas signal in TOPCAT suggests that in sicker patients with genuine structural HFpEF, aldosterone blockade was doing something real. The trial design just couldn’t isolate that.

Where This Leaves Spironolactone in Practice

Most HFpEF guidelines place spironolactone at Class IIb - may be considered, weak evidence. That classification is accurate, but it reflects a complicated evidentiary situation more than a true negative signal.

In practice, the reasoning most clinicians land on is roughly this: in HFpEF patients with volume overload, elevated filling pressures, echocardiographic features consistent with aldosterone-mediated remodeling, and no major contraindication, spironolactone is a defensible choice. The mechanism is sound, the drug is cheap, and the patient profile that showed benefit in the Americas analysis maps reasonably well to the patients you’re actually trying to treat.

What that reasoning requires is being explicit about what you’re doing - making a phenotype-based decision under evidentiary uncertainty - rather than treating TOPCAT as either a clean endorsement or a clean dismissal.

One thing that gets underweighted in the academic discussion is the real-world hyperkalemia problem. After RALES results were incorporated into guidelines and spironolactone prescribing expanded, Juurlink and colleagues documented a fourfold increase in serious hyperkalemia hospitalizations in the broader HF population. The RALES population was monitored carefully, used almost no ARBs (which weren’t yet widely available), and had fewer comorbidities than typical real-world patients. HFpEF patients - older, more CKD, more polypharmacy - sit squarely in the higher-risk group for this. That doesn’t change whether to use the drug, but it does change how closely you monitor and how conservative you are with dose escalation. In practice, many patients end up on 25 mg indefinitely, and that’s probably fine.

FINEARTS-HF and What It Actually Confirms

FINEARTS-HF enrolled 6,001 patients with HFmrEF or HFpEF and randomized them to finerenone - a non-steroidal, selective mineralocorticoid receptor antagonist - or placebo. After a median of 32 months, the primary composite of total worsening heart failure events plus cardiovascular death was reduced: RR 0.84 (95% CI 0.74–0.95), P=0.007.

The effect size isn’t large. But the trial is credible - adequate event rates, no obvious integrity issues, a population that actually had the disease - and the direction is consistent with what the Americas subgroup of TOPCAT was showing ten years earlier.

Hyperkalemia rates were higher than placebo (14.3% vs 6.9%), though manageable in the trial context. Finerenone’s selectivity also avoids the sex-hormone receptor off-target effects of spironolactone - gynecomastia, erectile dysfunction - that limit real-world tolerability, though a direct head-to-head comparison between agents wasn’t part of this trial.

What FINEARTS-HF does, taken together with everything else, is confirm that the mineralocorticoid receptor hypothesis in HFpEF was biologically correct. TOPCAT was asking the right question. The execution couldn’t answer it. FINEARTS-HF, with a better-tolerated agent and a cleaner trial, finally could.

Subscribe now

What This Changes About How I Think About Evidence in HFpEF

The arc from RALES to TOPCAT to FINEARTS-HF keeps pointing at the same problem: we design HFpEF trials around broadly defined populations to maximize enrollment, and then the biological signal we’re trying to detect gets averaged away across phenotypes where the therapy isn’t relevant.

A neutral trial in HFpEF is not automatically evidence of futility. Sometimes it’s evidence that the trial couldn’t answer the question - because of site integrity issues, because the population was too heterogeneous, because event rates were too low to detect a real effect in the patients who might benefit. Deciding which interpretation applies requires looking at the trial’s internal data, not just the headline result.

The harder problem is that we still can’t identify the “aldosterone-driven HFpEF phenotype” prospectively with enough precision to enrich for it in a trial - or, more practically, to be confident you’re selecting for it in clinic. E/e’, LV mass, natriuretic peptide levels, signs of volume overload - these are imperfect proxies. They’re what we have, and using them to guide a phenotype-based decision with spironolactone is reasonable. But it’s an inference, not a certainty.

TOPCAT didn’t tell us aldosterone blockade doesn’t work in HFpEF. It told us that running a large trial in an undifferentiated population, across sites with variable data integrity, produces an uninterpretable result. That’s a different lesson, and it’s one the field is still working through.

I’ve created a free one-page evidence framework to accompany this article. If you’d like a visual summary of the evidence, trial interpretation, and clinical takeaways, you can download the PDF below.

Aldosterone in HFpEF

—Cardio Concepts

Previous Ivabradine: When a Clean Mechanism Meets the Wrong Patient

Next Colchicine in Cardiology: An Old Drug Asking a New Question

References

1. Pitt B, Zannad F, Remme WJ, et al. The effect of spironolactone on morbidity and mortality in patients with severe heart failure. N Engl J Med. 1999;341(10):709–717. https://doi.org/10.1056/NEJM199909023411001

2. Pitt B, Pfeffer MA, Assmann SF, et al; TOPCAT Investigators. Spironolactone for heart failure with preserved ejection fraction. N Engl J Med. 2014;370(15):1383–1392. https://doi.org/10.1056/NEJMoa1313731

3. Pfeffer MA, Claggett B, Assmann SF, et al. Regional variation in patients and outcomes in the Treatment of Preserved Cardiac Function Heart Failure With an Aldosterone Antagonist (TOPCAT) trial. Circulation. 2015;131(1):34–42. https://doi.org/10.1161/CIRCULATIONAHA.114.013255

4. de Denus S, O’Meara E, Desai AS, et al. Spironolactone metabolites in TOPCAT — new insights into regional variation. N Engl J Med. 2017;376(17):1690–1692. https://doi.org/10.1056/NEJMc1612601

5. Edelmann F, Wachter R, Schmidt AG, et al; Aldo-DHF Investigators. Effect of spironolactone on diastolic function and exercise capacity in patients with heart failure with preserved ejection fraction: the Aldo-DHF randomized controlled trial. JAMA. 2013;309(8):781–791. https://doi.org/10.1001/jama.2013.905

6. Solomon SD, McMurray JJV, Vaduganathan M, et al; FINEARTS-HF Committees and Investigators. Finerenone in heart failure with mildly reduced or preserved ejection fraction. N Engl J Med. 2024;391(16):1475–1485. https://doi.org/10.1056/NEJMoa2407107

7. Juurlink DN, Mamdani MM, Lee DS, et al. Rates of hyperkalemia after publication of the Randomized Aldactone Evaluation Study. N Engl J Med. 2004;351(6):543–551. https://doi.org/10.1056/NEJMoa040135

Share

Leave a comment

Explore the Cardio Concepts Library

But that’s not what happened - or at least, not entirely. What actually happened is more complicated, and more revealing about how cardiology reasons about phenotype, mechanism, and who “counts” as a study population.

Before ACE Inhibitors, There Was This

1986. The RAAS blockade era hasn’t arrived yet. Heart failure management is digoxin and diuretics, with the implicit understanding that we’re managing symptoms rather than meaningfully changing natural history.

Then V-HeFT I comes out. Cohn and colleagues randomize 642 men - all VA hospital patients - to placebo, prazosin, or the combination of hydralazine 300 mg/day and isosorbide dinitrate 160 mg/day, on top of digoxin and diuretics. Two-year mortality: H-ISDN reduces it by 34% compared to placebo (P=0.028). Prazosin does nothing.

Worth noting: the overall follow-up mortality curve only reached borderline significance (P=0.053). The headline number - 34% at two years - was the primary endpoint, but it wasn’t a clean sweep across the entire follow-up period. That caveat matters when you’re reconstructing what people actually believed at the time.

Still, it was the first RCT to show a mortality signal in chronic heart failure. With a combination that had been around for decades.

The pharmacological rationale at the time was mostly hemodynamic: hydralazine reduces afterload, ISDN reduces preload. But hydralazine also inhibits superoxide production - it’s not just a vasodilator, it preserves NO from oxidative degradation. ISDN donates NO directly. Together they augment the NO-cGMP axis in a way that pure vasodilation doesn’t fully capture. That mechanistic layer became a lot more important later.

The Enalapril Problem

V-HeFT II, 1991. 804 men, enalapril 20 mg/day versus H-ISDN at the same doses. Mortality favors enalapril - roughly a 28% relative risk reduction at two years (P=0.016). Clear enough that ACE inhibition becomes standard and H-ISDN gets repositioned as the backup.

The story could have ended there. Old vasodilator combination, interesting historical footnote, superseded by better pharmacology.

Except the hemodynamic data doesn’t fit that narrative cleanly. H-ISDN produced larger improvements in LVEF and peak VO₂ than enalapril. Better functional capacity by those measures. A drug that lost on survival was doing something real physiologically - it just wasn’t converting into the mortality benefit that RAAS blockade achieved in this unselected population.

The question that wasn’t asked loudly enough: is this a drug problem, or a patient selection problem?

The Subgroup Nobody Acted On

Carson and colleagues published a post-hoc analysis of V-HeFT II in Journal of Cardiac Failure in 1999. When they stratified outcomes by race, Black patients showed similar mortality on H-ISDN and enalapril. The survival advantage driving the headline result came primarily from non-Black patients.

You have to be careful with post-hoc subgroup data, and the number of Black patients in V-HeFT II wasn’t large enough to draw firm conclusions. But the signal was internally consistent, the biological rationale was plausible, and it generated a testable hypothesis: maybe the outcome difference between H-ISDN and enalapril isn’t about which drug is globally superior. Maybe it’s about who the patient is.

Not race as biology. Race as a pointer - crude, contested, pointing at something that couldn’t yet be measured directly

A-HeFT: The Trial That Had to Stop Early

The African-American Heart Failure Trial tests that hypothesis prospectively. BiDil - fixed-dose hydralazine 37.5 mg plus ISDN 20 mg, three times daily - versus placebo in 1,050 self-identified Black patients with NYHA Class III–IV HFrEF (LVEF <35%), all on contemporary GDMT: ACE inhibitor or ARB, beta-blocker, aldosterone antagonist where appropriate.

The DSMB stops the trial in July 2004. Early termination.

All-cause mortality: 6.2% versus 10.2%. Hazard ratio 0.57 (95% CI 0.37–0.89). First heart failure hospitalization: HR 0.61. Quality of life: significantly better. Everything in the same direction, with effect sizes larger than most of what contemporary HFrEF trials deliver.

FDA approves BiDil in June 2005. First drug ever approved with a race-specific indication.

The approval is simultaneously reasonable and, in retrospect, a problem - not because the data is wrong, but because the trial design made it structurally impossible to answer the question it was supposed to answer.

There was no comparator group of non-Black patients. No internal control to determine whether the benefit was specific to Black patients or whether it was a benefit of H-ISDN in any patient with advanced heart failure, high oxidative burden, and inadequate NO signaling - regardless of race. You enrolled by self-identification, got approved by self-identification, and the mechanism question stayed open.

Early stopping also tends to inflate effect sizes. The 43% relative mortality reduction is probably real in direction, but the magnitude may be somewhat overstated.

What the Biology Actually Suggests

The mechanistic hypothesis behind BiDil is about nitric oxide bioavailability. There is evidence - some of it predating A-HeFT - that Black patients with heart failure have lower NO bioavailability and higher oxidative stress than matched non-Black patients. If that’s the actual driver, then the relevant phenotype isn’t race. It’s impaired NO-cGMP signaling and elevated redox burden. Race was an epidemiological shortcut - a way to enrich the trial for a biological state that couldn’t be measured at the bedside.

Cohn was explicit about this. The goal was never to define pharmacotherapy by race. It was to find a mechanism and a patient population. Race was the instrument available at the time; the biology underneath doesn’t sort by self-identification.

That reframing matters because it changes what A-HeFT actually proved. Not that H-ISDN is a Black patient drug. Rather, that H-ISDN dramatically reduces mortality in a population enriched for a specific pathophysiological state. The enrichment happened to use race. The mechanism is about redox-NO biology - which exists across populations, in varying degrees, for reasons that have more to do with disease severity, comorbidity burden, and access to care than with ancestry.

Subscribe now

What the Guidelines Do With This - and What They Don’t

The 2022 ACC/AHA/HFSA guidelines give H-ISDN a Class I, Level of Evidence A recommendation for self-identified Black patients with NYHA III–IV HFrEF on maximally tolerated GDMT. That’s the A-HeFT population replicated as an indication - defensible, since you indicate based on who was actually studied.

There’s also a Class IIa recommendation for patients who can’t tolerate RAAS inhibition or ARNI: bilateral renal artery stenosis, progressive renal dysfunction, recurrent angioedema. Mechanistically reasonable, and in practice underused in this group.

Both recommendations make sense. But there’s a gap that the guideline doesn’t address.

Take a non-Black patient - say, a middle-aged South Asian man with NYHA III HFrEF, already on full-dose sacubitril/valsartan and carvedilol, still with LVEF of 30% and persistent symptoms. The mechanism of H-ISDN should apply to him if his underlying biology involves the same NO-deficient, high-oxidative-stress phenotype. The guideline doesn’t say yes. It doesn’t say no either. It just doesn’t say anything, because he doesn’t fit A-HeFT’s enrollment criteria.

This is where race-as-proxy creates real clinical friction. We don’t have a validated, clinically accessible biomarker for NO bioavailability or oxidative stress - nothing you can order and use to phenotype patients at the bedside. So we’re left working with the enrollment instrument the trial used, applied to situations it wasn’t designed to answer.

Where This Actually Leaves You

The pragmatics are relatively clear. In a Black patient with advanced HFrEF symptomatic on optimal GDMT, H-ISDN has Level A evidence and effect sizes that are hard to argue with. Use it. The Class IIa indication for RAAS-intolerant patients is also underused - it’s a reasonable option for a group that otherwise has limited alternatives.

The harder question is the patient who sits outside the trial population but may have the same underlying biology. Here you’re extrapolating from mechanism rather than from enrollment criteria - which isn’t necessarily wrong, but you have to be honest that the evidence doesn’t extend there directly.

What A-HeFT ultimately showed is that there is a real subset of heart failure patients for whom NO augmentation, added to full neurohormonal blockade, dramatically changes outcomes. The trial found them through self-identified race. If the biology is what we think it is, that subset isn’t defined by race - it’s defined by a pathophysiological state that the trial happened to enrich for using the instruments available in 2001.

The prospective study that identifies these patients by mechanism rather than by race hasn’t been done. Until it is, the gap between what the evidence supports and what the biology predicts remains open. Recognizing that gap clearly is probably more useful than treating the 2004 approval as the final answer.

I've turned the central concept of this article into a one-page evidence framework. You can download it free below for future reference, teaching, or discussion.

The Race Proxy: H-ISDN, BiDil, and the NO-Deficiency Phenotype

—Cardio Concepts

Before Digoxin, Digitoxin, and the Question We Forgot to Ask

Next Ivabradine: When a Clean Mechanism Meets the Wrong Patient

References

1. Cohn JN, Archibald DG, Ziesche S, et al. Effect of vasodilator therapy on mortality in chronic congestive heart failure. N Engl J Med. 1986;314(24):1547–1552. https://doi.org/10.1056/NEJM198606123142404

2. Cohn JN, Johnson G, Ziesche S, et al. A comparison of enalapril with hydralazine–isosorbide dinitrate in the treatment of chronic congestive heart failure. N Engl J Med. 1991;325(5):303–310. https://doi.org/10.1056/NEJM199108013250502

3. Carson P, Ziesche S, Johnson G, Cohn JN. Racial differences in response to therapy for heart failure: analysis of the vasodilator-heart failure trials. J Card Fail. 1999;5(3):178–187.

4. Taylor AL, Ziesche S, Yancy C, et al; African-American Heart Failure Trial Investigators. Combination of isosorbide dinitrate and hydralazine in blacks with heart failure. N Engl J Med. 2004;351(20):2049–2057. https://doi.org/10.1056/NEJMoa042934

5. Heidenreich PA, Bozkurt B, Aguilar D, et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure. Circulation. 2022;145:e895–e1032. https://doi.org/10.1161/CIR.0000000000001063

   Share

   Leave a comment

   Explore the Cardio Concepts Library

That framing made sense in 1997. It makes less sense now.

The DIG Trial and What It Actually Said

When the DIG trial published in NEJM, it was immediately read as a verdict. Six thousand eight hundred patients. Digoxin versus placebo on top of ACE inhibitors and diuretics. Thirty-seven months of follow-up. All-cause mortality: 34.8% versus 35.1%. Risk ratio 0.99. P = 0.80.

Neutral. And the field moved on accordingly.

But the narrative that followed - digoxin doesn’t work, move on - skipped over a few things worth holding onto.

Hospitalization for heart failure dropped: from 34.7% to 26.8%. NNT around 13 over three years. In a disease where each hospitalization accelerates trajectory, that’s a clinically meaningful number. The patient population was predominantly NYHA I–II; only 2% were NYHA IV. The sickest patients - the ones most likely to be in our advanced HF clinics today, on maximally tolerated GDMT and still symptomatic - were barely in the room.

And most importantly: no beta-blockers in the background regimen. No MRAs as standard. No ARNIs, no SGLT2 inhibitors. None of it. A post-hoc analysis of DIG data later suggested that serum digoxin concentrations of 0.5–0.9 ng/mL were associated with better outcomes, while higher levels may have attenuated or reversed any benefit - but that was post-hoc, not prespecified, and it doesn’t translate cleanly into day-to-day dosing.

The conclusion that made sense in 1997 got locked in and transported forward, unchallenged, into an era where the patients were different, the background therapy was different, and the clinical question itself had shifted.

Why Digitoxin, and Why Now

Digoxin’s relationship with renal function is clinically inconvenient in ways that compound over time. In elderly HFrEF patients - which is most of the patients we’re talking about - renal function fluctuates. Dosing becomes uncertain. The therapeutic window penalizes variability.

Digitoxin is predominantly hepatically metabolized, making it substantially less renal-dependent than digoxin. Half-life around seven days. Serum concentrations more predictable across the kinds of patients who end up on it. The pharmacokinetics fit the population better.

The DIGIT-HF trial, published in NEJM in 2025, built its case around this. Fifty-five sites across Germany, Austria, and Serbia. 1,212 patients in the intention-to-treat analysis. HFrEF with LVEF ≤30% in NYHA II, or LVEF ≤40% in NYHA III–IV. At least six months of optimized GDMT including ARNI or ACEi, beta-blocker, MRA - and SGLT2 inhibitors in about 20%. Digitoxin titrated to a serum target of 8–18 ng/mL versus placebo. Median follow-up 36 months.

Primary composite - first HF hospitalization or all-cause death: HR 0.82 (95% CI 0.69–0.98). P = 0.03.

First positive cardiac glycoside trial in 28 years - though that framing, while accurate, can make the result sound cleaner than it is.

All-cause mortality alone: HR 0.86 - noninferiority established, superiority not reached. HF hospitalization alone: HR 0.85, directionally consistent but not significant. The trial was underpowered - it enrolled 1,240 of a planned 2,600 patients - which likely explains why the individual components didn’t hold on their own.

There’s also a statistical issue that deserves honest acknowledgment. A preprint analysis identified a violation of the proportional hazards assumption (P = 0.019): the relative benefit between groups wasn’t constant across follow-up time, appearing more pronounced earlier. This doesn’t overturn the composite result, but it means the single HR may be averaging across a more complex time-dependent effect. The trial is positive, but it’s modestly positive, in a narrower population than DIG, with caveats worth keeping in mind.

AF as the Phenotype

A post-hoc analysis from DIGIT-HF, published in European Heart Journal in 2026, stratified the 1,212 patients by AF at baseline. About 27% - 330 patients - had AF.

The interaction term for mortality was significant: P = 0.01.

In patients with AF: HR for all-cause mortality 0.55 (95% CI 0.35–0.92). Estimated NNT for death over 36 months: approximately 7. In patients without AF: HR 1.01 (0.78–1.30).

An estimated NNT of 7 for a mortality endpoint stops you, even when you know the caveats - not prespecified, small subgroup, the proportional hazards issue from the main trial carries over here. But the signal has mechanistic coherence, and that matters when you’re deciding how seriously to take it.

In AF + HFrEF, there’s a specific hemodynamic problem that doesn’t get discussed enough. The ventricular rate in AF is a driver of ongoing LV dysfunction - not in the pure tachycardia-mediated cardiomyopathy sense, but in the sense that uncontrolled rate in an already failing ventricle worsens filling, increases oxygen demand, and shortens diastole at exactly the wrong moment. You want rate control. The problem is how you get it.

Beta-blockers are the default. But in advanced HFrEF with low output and elevated filling pressures, the dose that achieves adequate rate control may be the dose that tips the hemodynamics. This is a tension that comes up in practice more than the guidelines acknowledge - the patient who most needs rate control is often the one least able to tolerate aggressive beta-blockade.

Digitoxin works differently here. The ventricular rate slows through vagal augmentation at the AV node, not through beta-adrenergic blockade. And simultaneously, the positive inotropic effect is supporting a ventricle that is both rate-burdened and contractility-limited. When you see that mortality signal in the AF subgroup, the biology behind it isn’t arbitrary - it’s a drug doing two things at once in a context where both things matter.

Where CASTLE-AF and EAST-AFNET 4 Fit

It’s worth being clear about where rhythm control evidence sits, because it shapes how you think about the digitoxin niche.

CASTLE-AF (NEJM, 2018) - catheter ablation versus rate/rhythm control in HFrEF + AF - showed HR 0.62 for the composite of death and HF hospitalization. EAST-AFNET 4 (NEJM, 2020) extended the argument into earlier AF: early rhythm control in newly diagnosed AF reduced CV death, stroke, and HF hospitalization versus usual care (HR 0.79).

The direction is consistent: if you can restore and maintain sinus rhythm in a patient with structural heart disease, it probably matters. Rate control is increasingly a fallback, not a primary strategy - and the EAST-AFNET 4 data suggest that even waiting too long to pursue rhythm control has consequences.

Which is exactly why the digitoxin question is specific. It’s not competing with ablation when ablation is feasible. It’s not a first-line rate control agent. The context is the patient with established HFrEF, persistent or permanent AF, where rhythm control isn’t happening or hasn’t worked, already on maximally tolerated GDMT, and where beta-blocker titration keeps running into hemodynamic limits. That patient is real and shows up regularly in advanced HF practice.

Subscribe now

What I’m Actually Thinking About in Practice

DIGIT-HF doesn’t change things overnight. The trial was underpowered. The AF subgroup analysis wasn’t prespecified. SGLT2 inhibitor use was only 20%, which doesn’t fully reflect what contemporary GDMT looks like in centers with real HF infrastructure. And digitoxin has no approval pathway in most markets outside central Europe.

But the question it reopens is worth taking seriously.

For years the implicit logic was: DIG showed no mortality benefit, so cardiac glycosides are a relic, and we use them reluctantly when symptom control demands it. That logic assumed the DIG population, the DIG background therapy, and the DIG clinical context were still the relevant frame of reference.

They’re not. In a patient with advanced HFrEF, AF, on optimized GDMT, where beta-blocker titration is hitting a wall - the question of whether digitoxin has a role is a real clinical question, not a historical one. The AF subgroup signal is the most specific piece of evidence we’ve had pointing toward a defined phenotype where the benefit may be concentrated.

An estimated NNT of 7 for mortality, post-hoc or not, earns a prospective test. Whether that trial gets funded is a separate question. But the hypothesis now has mechanistic rationale, pharmacokinetic justification, and outcome data pointing in the same direction.

That’s more than it had in 2024.

A one-page framework summarizing the key takeaways from DIG, DIGIT-HF, and the evolving role of cardiac glycosides in HFrEF. Download below

Digoxin is dead?

—Cardio Concepts

Next The Race Question We Never Fully Answered

References

1. The Digitalis Investigation Group. The effect of digoxin on mortality and morbidity in patients with heart failure. N Engl J Med. 1997;336(8):525–533. https://doi.org/10.1056/NEJM199702203360801

2. Bavendiek U, et al.; DIGIT-HF Study Group. Digitoxin in patients with heart failure and reduced ejection fraction. N Engl J Med. 2025;393(12):1155–1165. https://doi.org/10.1056/NEJMoa2415471

3. Bavendiek U, Thomas NH, et al.; DIGIT-HF Study Group. Digitoxin in patients with heart failure and atrial fibrillation: post-hoc analysis of the DIGIT-HF trial. Eur Heart J. 2026. https://doi.org/10.1093/eurheartj/ehag379

4. Marrouche NF, et al.; CASTLE-AF Investigators. Catheter ablation for atrial fibrillation with heart failure. N Engl J Med. 2018;378(5):417–427. https://doi.org/10.1056/NEJMoa1707855

5. Kirchhof P, et al.; EAST-AFNET 4 Trial Investigators. Early rhythm-control therapy in patients with atrial fibrillation. N Engl J Med. 2020;383(14):1305–1316. https://doi.org/10.1056/NEJMoa2019422

   Share

   Leave a comment

   Explore the Cardio Concepts Library

And yet, I keep finding myself drawn back to them.

Partly because medicine has a habit of revisiting ideas it once abandoned. Partly because many of these drugs seem to be returning - not unchanged, but reinterpreted through the lens of new biology, new patient phenotypes, and new clinical questions.

Over the years, I’ve noticed something interesting: a negative trial rarely means a mechanism is completely wrong. More often, it means we asked the wrong question, studied the wrong population, measured the wrong outcome, or intervened at the wrong moment in the disease process.

That’s what this series is about.

Not nostalgia. Not an attempt to revive outdated therapies. But a closer look at several “old” cardiovascular drugs that have unexpectedly re-entered modern discussions: digitalis, hydralazine–isosorbide dinitrate, ivabradine, spironolactone, and colchicine.

Each of them carries a story that is more complicated than the headline most of us remember. Some were overshadowed by newer therapies. Some were limited by flawed trials. Some turned out to be useful only in very specific phenotypes. And some are helping us understand cardiovascular disease in ways their original developers could never have imagined.

What interests me most is not whether these drugs are old or new, it’s how their stories remind us that medicine rarely moves in a straight line.

Sometimes progress comes from discovering something new, and sometimes it comes from asking better questions about something we thought we already understood.

This series is my attempt to revisit a few of those questions.

In the weeks ahead, we’ll revisit five stories:

• Digoxin, Digitoxin, and the Question We Forgot to Ask

• The Race Question We Never Fully Answered: H-ISDN and the Biology Behind BiDil

• Ivabradine: When a Clean Mechanism Meets the Wrong Patient

• The TOPCAT Problem: What a Flawed Trial Taught Us About HFpEF, Aldosterone, and the Limits of Evidence

• Colchicine in Cardiology: An Old Drug Asking a New Question

And finally, a synthesis piece:

• Old Drugs, New Questions: Five Stories About What We Got Wrong the First Time

Each article stands on its own. Together, they tell a larger story about how cardiovascular medicine evolves - not only through new discoveries, but also through revisiting old ideas with better questions.

—Cardio Concepts

Subscribe now

Explore the Cardio Concepts Library

There’s something quietly counterintuitive about the way we treat heart failure today. We give drugs that lower blood pressure to patients who are already hypoperfused. We block the very hormones that are trying to keep their cardiac output up. And somehow - consistently, across decades of trials - this makes them live longer.

If you only have the hemodynamic model in your head, none of this makes sense.

The hemodynamic model says: heart is weak → pump harder. It’s the obvious logic. When output falls, recruit compensatory mechanisms. Boost contractility. Hold onto sodium and water. Constrict the periphery. Keep perfusion pressure up. For much of the 20th century, this was the mental framework guiding treatment - and it led, with remarkable directness, to a class of drugs called inotropes.

And then the data started coming in, and the model started to fall apart.

The Counterpoint That Changed Everything

In 1991, the PROMISE trial enrolled over a thousand patients with severe chronic heart failure and randomized them to milrinone - an oral phosphodiesterase inhibitor that improves both contractility and vasodilation. On paper, it does exactly what you’d want: it raises cardiac output, reduces filling pressures, and makes patients feel better in the short term. The hemodynamics are unambiguously improved.

Mortality in the milrinone group increased by 28%.

This is the kind of finding that doesn’t just change a treatment recommendation. It changes a question. If improving hemodynamics makes patients die faster, then hemodynamics isn’t really what’s killing them. So what is?

The question sat there, partly answered by trials that had already been published. V-HeFT I, from 1986, had shown that pure vasodilation - hydralazine combined with isosorbide dinitrate - reduced two-year mortality from 34.3% to 25.6% compared to placebo. Not dramatic by modern standards, but notable for what it implied: offloading the heart from the outside, without touching contractility, was already doing something right. The mechanism wasn’t clear yet. The concept was just beginning to take shape.

Then came CONSENSUS, in 1987. Enalapril in patients with NYHA class IV heart failure. Six-month mortality dropped from 44% to 26% - a 40% relative risk reduction. Striking enough. But the detail that mattered most to the investigators was where the benefit came from: entirely from reductions in progressive pump failure deaths, not sudden death. This wasn’t just symptom relief. Something was slowing the disease itself.

SOLVD-Treatment, in 1991, confirmed and broadened the picture. In the far more common NYHA II–III population, enalapril reduced all-cause mortality by 16% and heart failure hospitalizations from 36.6% to 25.8%. The signal was clear, consistent, and durable.

The question was: why? What was enalapril actually doing that made patients live longer, when milrinone - which improved the same hemodynamic parameters - killed them?

The Hidden Concept

Milton Packer’s 1992 paper in the Journal of the American College of Cardiology wasn’t a clinical trial. It was a conceptual argument - an attempt to name what the evidence was already showing. He called it the neurohormonal hypothesis.

The core idea was this: heart failure doesn’t kill primarily because cardiac output is low. It kills because of what happens next. When output falls, the body activates two ancient survival systems - the sympathetic nervous system and the renin-angiotensin-aldosterone axis. These systems evolved for hemorrhage and dehydration. They constrict blood vessels, retain sodium, accelerate the heart. Short-term, they’re life-saving.

Long-term, in a patient whose heart is already struggling, they become relentless destructive forces. Sustained vasoconstriction increases wall stress on a ventricle that cannot compensate. Neurohormonal signaling drives cardiomyocyte apoptosis. The electrical substrate becomes increasingly unstable. The heart remodels - dilates, thins, loses its geometry - and its function deteriorates further, which activates the neurohormonal system further still. A loop with no natural exit.

In this framework, the reason ACE inhibitors work isn’t that they lower blood pressure. It’s that they interrupt the loop. They cut the neurohormonal cascade at one of its key amplification steps, relieving the myocardium of a sustained molecular assault it can no longer survive. And the reason milrinone fails - despite doing the hemodynamic equivalent of what ACE inhibitors do - is that it doesn’t touch the cascade at all. It borrows cardiac reserve while the loop keeps running.

This distinction is what separates afterload reduction in its modern sense from vasodilation in its older sense. The target isn’t the blood vessel. It’s the signaling system behind the blood vessel.

The Concept Evolves

If the neurohormonal hypothesis explained why ACE inhibitors worked, it also implied a question: what if you could do better than blocking one limb of the system?

PARADIGM-HF, published in 2014, was built around exactly that question. Sacubitril/valsartan - an angiotensin receptor-neprilysin inhibitor - combines blockade of the RAAS axis with preservation of natriuretic peptides, the heart’s own endogenous vasodilators. Rather than simply suppressing the vasoconstrictor arm, it simultaneously amplifies the vasodilator arm. Two axes, rebalanced.

Compared to enalapril alone, sacubitril/valsartan reduced the primary composite endpoint by 20% (HR 0.80, P<0.001), cardiovascular mortality by 20%, and all-cause mortality by 16%. These are not marginal gains on top of an established therapy. They represent a meaningful conceptual step: the recognition that neurohormonal unloading isn’t a single lever, but a system in which both sides matter.

From a clinical standpoint, PARADIGM-HF didn’t just add a drug. It sharpened the concept. Optimal afterload reduction in HFrEF is not about achieving a target blood pressure. It’s about restoring hormonal balance in a system that has been chronically dysregulated.

Where the Concept Gets More Complicated

Here is where the story becomes honest about its limits.

When the same conceptual framework was applied to heart failure with preserved ejection fraction - and it was, systematically, in PARAGON-HF - the results were sobering. Sacubitril/valsartan in HFpEF reduced the primary endpoint with a rate ratio of 0.87, just crossing the conventional boundary of statistical significance at P=0.06. Not quite there.

The instinct might be to call this a trial failure. But it’s more useful to call it a signal that afterload in HFpEF is a different problem - physically, not just semantically.

In HFrEF, the dominant afterload component is steady-state: systemic vascular resistance, the persistent force against which the heart ejects. This is the target of ACE inhibitors, ARBs, and ARNI. It’s the pressure you can read on a cuff.

In HFpEF, the more pathologically relevant component is pulsatile - the oscillating load driven by aortic stiffness and wave reflection. As the aorta ages and stiffens, the pressure wave that travels down into the arterial tree reflects back earlier and earlier in the cardiac cycle, arriving in late systole rather than diastole. The result: the left ventricle faces an additional, late-arriving wall stress burden precisely when it should be beginning to relax. The myocardium cannot relax efficiently. Diastolic function deteriorates.

Reddy and colleagues demonstrated this concretely in 2017, using invasive hemodynamics and arterial tonometry in patients with HFpEF versus matched controls. The pulsatile aortic load in HFpEF rose dramatically with exercise - independently of mean blood pressure. Two patients with identical blood pressures on a cuff can have profoundly different pulsatile afterloads. Weber and Chirinos, in a 2018 state-of-the-art review, systematized the framework: aortic stiffness, wave reflection, and late-systolic loading are distinct hemodynamic targets that conventional antihypertensive strategies don’t adequately address.

This is, in part, why HFpEF has been so resistant to the treatment strategies that transformed HFrEF. The problem isn’t the same problem, even when the blood pressure is similar. The concept of afterload reduction had to grow again - from neurohormonal unloading to vascular phenotyping - to even begin framing the right questions.

A Boundary Worth Remembering

There is one more layer to this, and it keeps the concept honest.

The EPHESUS post-hoc analysis, published in 2020, examined the relationship between diastolic blood pressure and outcomes in patients with left ventricular dysfunction after myocardial infarction. In patients who had not undergone revascularization, diastolic blood pressure below 70 mmHg was associated with a markedly increased risk of death (adjusted HR 1.80, P<0.001). In patients who had been revascularized, the association disappeared entirely.

The physiological explanation is not subtle. Coronary perfusion of the subendocardium occurs primarily during diastole, and in a ventricle with elevated end-diastolic pressure and impaired relaxation, the perfusion gradient is already narrowed. Aggressive afterload reduction in a patient with residual epicardial disease can tip that balance - improving loading conditions while simultaneously reducing the driving pressure for myocardial perfusion.

Unloading is not unconditionally beneficial. The therapeutic window is real.

Subscribe now

What This All Adds Up To

Over nearly four decades, the concept of afterload reduction has traveled a considerable distance. It began as a hemodynamic intervention - lower the resistance, help the heart eject. It became, through the neurohormonal hypothesis, an understanding of disease mechanism - interrupt the cascade that is slowly destroying the myocardium. It evolved further, through PARADIGM-HF, into a concept of biological rebalancing - restore equilibrium between the vasoconstrictor and vasodilator arms of the neurohormonal system. And it now faces its most complex iteration in HFpEF, where the relevant afterload is not the pressure on a cuff but the stiffness of an aorta and the timing of a reflected wave.

None of these framings made the previous ones obsolete. They layered. What changed, at each step, was not the target organ but the depth of the question being asked about it.

And maybe that’s the real lesson embedded in this body of evidence: in heart failure, the most important things happening are not the things you can measure at the bedside. They’re happening at the level of receptors and signaling cascades, in the elastic properties of arterial walls, in the timing of pressure waveforms that no blood pressure cuff captures.

Treating afterload means understanding what’s actually loading the heart - which turns out to be a more complicated question than it first appeared.

We often treat afterload as a number. This framework explores how it gradually became a biological concept instead. Free PDF below.

AFTERLOAD REDUCTION

—Cardio Concepts

References

1. Cohn JN, Archibald DG, Ziesche S, Franciosa JA, Harston WE, Tristani FE, Dunkman WB, Jacobs W, Francis GS, Flohr KH, et al. Effect of vasodilator therapy on mortality in chronic congestive heart failure: results of a Veterans Administration Cooperative Study. N Engl J Med. 1986;314(24):1547–1552. doi:10.1056/NEJM198606123142404

2. CONSENSUS Trial Study Group. Effects of enalapril on mortality in severe congestive heart failure: results of the Cooperative North Scandinavian Enalapril Survival Study (CONSENSUS). N Engl J Med. 1987;316(23):1429–1435. doi:10.1056/NEJM198706043162301

3. The SOLVD Investigators (Yusuf S, Pitt B, Davis CE, Hood WB Jr, Cohn JN). Effect of enalapril on survival in patients with reduced left ventricular ejection fractions and congestive heart failure. N Engl J Med. 1991;325(5):293–302. doi:10.1056/NEJM199108013250501

4. Packer M, Carver JR, Rodeheffer RJ, Ivanhoe RJ, DiBianco R, Zeldis SM, Hendrix GH, Bommer WJ, Elkayam U, Kukin ML, et al; PROMISE Study Research Group. Effect of oral milrinone on mortality in severe chronic heart failure. N Engl J Med. 1991;325(21):1468–1475. doi:10.1056/NEJM199111213252103

5. Packer M. The neurohormonal hypothesis: a theory to explain the mechanism of disease progression in heart failure. J Am Coll Cardiol. 1992;20(1):248–254. doi:10.1016/0735-1097(92)90167-L

6. McMurray JJV, Packer M, Desai AS, Gong J, Lefkowitz MP, Rizkala AR, Rouleau JL, Shi VC, Solomon SD, Swedberg K, Zile MR; PARADIGM-HF Investigators and Committees. Angiotensin–neprilysin inhibition versus enalapril in heart failure. N Engl J Med. 2014;371(11):993–1004. doi:10.1056/NEJMoa1409077

7. Reddy YNV, Andersen MJ, Obokata M, Koepp KE, Kane GC, Melenovsky V, Olson TP, Borlaug BA. Arterial stiffening with exercise in patients with heart failure and preserved ejection fraction. J Am Coll Cardiol. 2017;70(2):136–148. doi:10.1016/j.jacc.2017.05.029

8. Weber T, Chirinos JA. Pulsatile arterial haemodynamics in heart failure. Eur Heart J. 2018;39(43):3847–3854. doi:10.1093/eurheartj/ehy346

9. Solomon SD, McMurray JJV, Anand IS, Ge J, Lam CSP, Maggioni AP, Martinez F, Packer M, Pfeffer MA, Pieske B, et al; PARAGON-HF Investigators and Committees. Angiotensin–neprilysin inhibition in heart failure with preserved ejection fraction. N Engl J Med. 2019;381(17):1609–1620. doi:10.1056/NEJMoa1908655

10. Böhm M, Ferreira JP, Mahfoud F, Duarte K, Pitt B, Zannad F, Rossignol P. Myocardial reperfusion reverses the J-curve association of cardiovascular risk and diastolic blood pressure in patients with left ventricular dysfunction and heart failure after myocardial infarction: insights from the EPHESUS trial. Eur Heart J. 2020;41(17):1673–1683. doi:10.1093/eurheartj/ehaa132

    Share

    Leave a comment

    Explore the Cardio Concepts Library

If I had to distill this entire seven-part series into a single sentence, it would be this: it took us roughly two decades to recognize that atrial fibrillation is not a homogeneous disease, and anticoagulation is not a binary decision.

The journey moved through landmark trials, unexpected results, moments where we had to go back and question what we thought we knew - and a few failures that ended up being more useful than some of the successes.

Where the story starts

The first piece in the series Why Anticoagulation Is No Longer One-Size-Fits-All asked the foundational question: why is anticoagulation no longer one-size-fits-all? Not because we have more drugs to choose from. But because we started looking at patients differently.

From ESC 2020 to ACC/AHA 2023 to ESC 2024, the major guidelines weren’t just updating recommendations - they were reflecting a genuine shift in treatment philosophy. Lip’s ABC pathway, the 4S-AF descriptive model, ESC 2024’s decision to remove female sex from the OAC initiation threshold in CHA₂DS₂-VA - all of it points in the same direction: decisions in AF need to start from patient phenotype, not from a score cutoff.

A paradox partially resolved

For a long time, the bleeding–thrombosis paradox The Bleeding–Thrombosis Paradox was the practical ceiling on how aggressively clinicians would anticoagulate.

In the warfarin era, a high HAS-BLED score often meant no anticoagulation. Clinically, that felt like caution. But Friberg’s Swedish national cohort of 182,678 AF patients showed why it was systematically wrong: net clinical benefit of OAC remained positive across virtually all HAS-BLED strata, because the same risk factors that predicted bleeding also predicted stroke. Withholding OAC from the highest-risk patients meant withholding it from the people who needed it most.

The four pivotal trials - RE-LY, ROCKET-AF, ARISTOTLE, ENGAGE AF-TIMI 48 - and Ruff’s 2014 meta-analysis (n=71,683) then put numbers to what NOACs actually changed: ICH reduced by 52% (RR 0.48), hemorrhagic stroke by 51% (RR 0.49), all-cause mortality by 10% (RR 0.90). GI bleeding increased modestly with standard-dose NOACs - that trade-off didn’t disappear. But ICH, which is the bleeding outcome that actually kills people or leaves them disabled, dropped by half. ARISTOTLE with apixaban was the only trial to come out superior to warfarin on stroke, major bleeding, and mortality at the same time.

When frailty changes the equation

The frailty piece Frailty Changes Everything is probably the most important for day-to-day practice, and also the one most at risk of being misread.

Elderly, frail patients prone to falls - this is the group we worry most about when prescribing OAC, and also the group with the highest stroke risk. The prespecified age subgroup from ENGAGE AF-TIMI 48 showed something that cuts against clinical instinct: the absolute benefit of edoxaban increased with age. In patients ≥75, edoxaban prevented 82 major bleeding events and 73 ICH events per 10,000 patient-years compared to warfarin. The older the patient, the more there was to gain.

ELDERCARE-AF then opened the specific question of patients too frail or too high-risk for standard NOAC doses - edoxaban 15 mg/day versus placebo in Japanese patients aged ≥80 (mean age 86.6, mean weight 50.6 kg). Stroke was reduced by 66% relative (HR 0.34). But the context matters: placebo-controlled, low-body-weight Asian population. How applicable it is outside that setting is a real question, not a formality.

FRAIL-AF delivered a different kind of lesson. Frail patients already stable on VKA were randomized to switch to NOAC or continue - and switching increased bleeding significantly (HR 1.69). The instinct is to read this as “NOACs are more dangerous in frail patients.” That’s not what the trial asked. It asked whether switching someone who’s already doing well is a good idea. The answer is no. Whether NOAC or VKA is better as a starting choice in a treatment-naive frail patient is a separate question, and one that remains genuinely unanswered.

When antiplatelets collide with anticoagulation

AF plus a recent stent is one of the most common genuine dilemmas in cardiology practice When Two Risks Collide, and this is an area where the evidence is both relatively solid and has clearly changed what people actually do.

PIONEER AF-PCI, RE-DUAL PCI, AUGUSTUS, and ENTRUST-AF PCI all arrived between 2016 and 2019, followed by AFIRE and Lopes’ network meta-analysis. The consistent signal across all of them: NOAC plus P2Y12 inhibitor, drop aspirin early, is the optimal regimen after PCI in AF patients.

AUGUSTUS was the trial that settled the mechanism. Its 2×2 factorial design let investigators separate two decisions simultaneously - which OAC to use, and whether to include aspirin. The aspirin result was stark: adding it nearly doubled major and clinically relevant bleeding (HR 1.89) without improving ischemic outcomes. AFIRE then extended the logic to the stable phase: in AF patients with CAD more than 12 months out from revascularization, NOAC monotherapy was non-inferior on efficacy and actually superior on safety and mortality compared to NOAC plus antiplatelet.

The largest evidence gap: CKD

This CKD and Anticoagulation is the most uncomfortable piece in the series, because it’s mostly about what we don’t know.

In mild-to-moderate CKD, NOACs perform well. The prespecified CKD 4 subgroup from ARISTOTLE (CrCl 25–30 mL/min) showed apixaban reduced major bleeding with HR 0.34, and PK data confirmed no problematic drug accumulation. But as kidney function declines toward ESKD and hemodialysis, the evidence base essentially collapses.

The three RCTs in this space - RENAL-AF, AXADIA-AFNET 8, and Valkyrie - were all severely underpowered. RENAL-AF enrolled only 154 of its 760-patient target (~20%) before stopping due to enrollment failure, not any clinical signal. The confidence interval for the primary endpoint ran from 0.63 to 2.30 - it can’t rule out either meaningful benefit or meaningful harm. AXADIA used apixaban 2.5 mg BID for all hemodialysis patients, which is a questionable design choice: Siontis’ observational cohort of 25,523 ESKD patients suggests it’s the 5 mg dose - not 2.5 mg - that reduces stroke and mortality in this population. Valkyrie had the most positive signal for rivaroxaban but with n=132 and warfarin TTR of 50.7% in the control arm, making it hard to know how much of the apparent benefit is drug performance versus poorly managed VKA.

In ESKD plus AF, we are making decisions based on evidence that isn’t strong enough to confidently support any specific recommendation. That’s uncomfortable. It’s also just where we are, and probably worth being explicit about rather than reaching for the least-underpowered trial and treating it as an answer.

Residual risk and the device question

Even when patients are on the right NOAC at the right dose, some still have strokes When the Drug Is Right but the Risk Isn't Gone. That’s a problem the field has only recently started naming explicitly.

ORBIT-AF II (n=5,738) quantified one part of it: in real-world practice, 9.4% of patients were underdosed and 3.4% overdosed. Underdosing increased CV mortality (aHR 1.91) without preventing stroke; overdosing increased major bleeding (aHR 2.17). A meaningful share of what we call “residual risk” is actually just prescribing error.

But there’s also a genuinely biological piece. Seiffge’s pooled analysis of breakthrough strokes - ischemic strokes in AF patients already on OAC - found a recurrence rate of 7.3% per year, higher than the general AF population. Switching OAC type after a breakthrough stroke didn’t reduce recurrence meaningfully (aHR 1.2). This is observational data with real confounding limitations, so the conclusions have to stay tentative. But it raises a question worth sitting with: in some AF patients, is the thrombotic mechanism not fully addressed by FXa or thrombin inhibition alone?

COMPASS added another angle - rivaroxaban 2.5 mg BID plus aspirin in stable atherosclerotic disease (no AF) reduced the composite of stroke, MI, and CV death (HR 0.76) versus aspirin alone, which demonstrated that residual atherothrombotic risk operates through a different pathway and responds to a different intervention. And at the device end of the spectrum, PRAGUE-17 and OPTION provided evidence for LAA closure as a legitimate alternative in patients where high thrombotic and high bleeding risk make long-term anticoagulation genuinely difficult to manage.

Can we uncouple thrombosis from bleeding?

The final piece The Cardiology of Wishful Thinking is about where the field is heading - and what recent phase 3 results tell us about how complicated that question actually is.

The premise behind FXIa inhibition is that FXI/XIa is disproportionately important for pathological clot formation relative to physiological hemostasis. Block it, and you should be able to reduce thrombosis without impairing normal bleeding control. Phase 2 data with asundexian in AF (PACIFIC-AF) looked consistent with that idea.

OCEANIC-AF (phase 3, n=14,810) then answered the efficacy question directly: stroke or systemic embolism occurred at nearly four times the rate with asundexian compared to apixaban (HR 3.79), despite a clear reduction in bleeding. Stopped early by the DSMB. This is a specific failure, not a class failure - it tells us that left atrial thrombus in AF depends less on the FXIa pathway than venous or atherothrombotic clot does. Whether that’s a problem with FXIa inhibition in general, or with the degree of inhibition asundexian achieves, or something more fundamental about LAA thrombus biology, is still being worked out.

Abelacimab approaches the coagulation cascade at a different point: rather than blocking FXIa after activation, it locks FXI in its zymogen form and prevents activation from occurring. In AZALEA-TIMI 71 (n=1,287), abelacimab at 150 mg monthly reduced major and clinically relevant bleeding by 62% versus rivaroxaban (HR 0.38), with zero ICH events in the abelacimab arms. Phase 2b only - efficacy conclusions aren’t possible yet. Phase 3 (LILAC-TIMI 76) is enrolling.

OCEANIC-STROKE, reported at ISC 2026, gave the class its first phase 3 success: asundexian plus antiplatelet reduced recurrent ischemic stroke by approximately 26% (csHR ~0.74) without increasing major bleeding. That matters - but it’s in non-cardioembolic stroke, where the thrombotic mechanism relies more heavily on contact pathway activation than it does in AF. The emerging picture is that FXIa inhibition works where the biology calls for it, and doesn’t where it doesn’t.

The framework behind the series

Looking back, the series has been building toward a single practical reorientation.

We spent years asking “should we anticoagulate?” Then the question shifted to “which anticoagulant?” The more important shift - the one these seven posts have tried to make explicit - is learning to ask something more specific first: what is this patient’s clinical phenotype, and what does the evidence actually say for that phenotype?

An 85-year-old on hemodialysis is not the same clinical problem as a 60-year-old with paroxysmal AF and no comorbidities. A patient three days out from coronary stenting is not the same as someone with stable CAD managed for two years. A frail patient who is already stable and well-controlled on warfarin is not the same as a frail patient starting OAC for the first time. These distinctions matter - not just in theory but in what the evidence actually supports, including where it runs out.

That’s the framework I’ve put together in the companion resources attached to this post: a free PDF summary of the full series, and an editable PPTX with full speaker notes structured around the seven phenotype categories. Both available at the link below.

NOAC Slide Deck Free PDF

NOAC Slide Deck (PPTX)

—Cardio Concepts

Thanks to everyone who read through this from the beginning. It was never meant to be easy reading, but I hope the reasoning behind the evidence felt a little more visible by the end.

References

1. Hindricks G, et al. 2020 ESC Guidelines for the diagnosis and management of atrial fibrillation. Eur Heart J. 2021;42(5):373–498. https://doi.org/10.1093/eurheartj/ehaa612

2. Joglar JA, et al. 2023 ACC/AHA/ACCP/HRS Guideline for Diagnosis and Management of Atrial Fibrillation. Circulation. 2024;149(1):e1–e156. https://doi.org/10.1161/CIR.0000000000001193

3. Van Gelder IC, et al. 2024 ESC Guidelines for the management of atrial fibrillation. Eur Heart J. 2024;45(36):3314–3414. https://doi.org/10.1093/eurheartj/ehae176

4. Lip GYH. The ABC pathway: an integrated approach to improve AF management. Nat Rev Cardiol. 2017;14(11):627–628. https://doi.org/10.1038/nrcardio.2017.153

5. Ruff CT, et al. Comparison of the efficacy and safety of new oral anticoagulants with warfarin in patients with atrial fibrillation: a meta-analysis of randomised trials. Lancet. 2014;383(9921):955–962. https://doi.org/10.1016/S0140-6736(13)62343-0

6. Connolly SJ, et al. Dabigatran versus Warfarin in Patients with Atrial Fibrillation. N Engl J Med. 2009;361(12):1139–1151. https://doi.org/10.1056/NEJMoa0905561

7. Patel MR, et al. Rivaroxaban versus Warfarin in Nonvalvular Atrial Fibrillation. N Engl J Med. 2011;365(10):883–891. https://doi.org/10.1056/NEJMoa1009638

8. Granger CB, et al. Apixaban versus Warfarin in Patients with Atrial Fibrillation. N Engl J Med. 2011;365(11):981–992. https://doi.org/10.1056/NEJMoa1107039

9. Giugliano RP, et al. Edoxaban versus Warfarin in Patients with Atrial Fibrillation. N Engl J Med. 2013;369(22):2093–2104. https://doi.org/10.1056/NEJMoa1310907

10. Friberg L, Rosenqvist M, Lip GYH. Net clinical benefit of warfarin in patients with atrial fibrillation. Circulation. 2012;125(19):2298–2307. https://doi.org/10.1161/CIRCULATIONAHA.111.055079

11. Carnicelli AP, et al. Direct Oral Anticoagulants Versus Warfarin in Patients with Atrial Fibrillation: Patient-Level Network Meta-Analyses. Circulation. 2022;145(4):242–255. https://doi.org/10.1161/CIRCULATIONAHA.121.056355

12. Okumura K, et al. Low-Dose Edoxaban in Very Elderly Patients with Atrial Fibrillation. N Engl J Med. 2020;383(18):1735–1745. https://doi.org/10.1056/NEJMoa2012883

13. Kato ET, et al. Efficacy and Safety of Edoxaban in Elderly Patients With Atrial Fibrillation in the ENGAGE AF-TIMI 48 Trial. J Am Heart Assoc. 2016;5(5):e003432. https://doi.org/10.1161/JAHA.116.003432

14. Joosten LPT, et al. Safety of Switching From a Vitamin K Antagonist to a Non–Vitamin K Antagonist Oral Anticoagulant in Frail Older Patients With Atrial Fibrillation: FRAIL-AF. Circulation. 2024;149(4):279–289. https://doi.org/10.1161/CIRCULATIONAHA.123.066485

15. Savelieva I, et al. EHRA expert consensus document on the management of arrhythmias in frailty syndrome. Europace. 2023;25(4):1249–1276. https://doi.org/10.1093/europace/euac123

16. Gibson CM, et al. Prevention of Bleeding in Patients with Atrial Fibrillation Undergoing PCI. N Engl J Med. 2016;375(25):2423–2434. https://doi.org/10.1056/NEJMoa1611594

17. Cannon CP, et al. Dual Antithrombotic Therapy with Dabigatran after PCI in Atrial Fibrillation. N Engl J Med. 2017;377(16):1513–1524. https://doi.org/10.1056/NEJMoa1708454

18. Lopes RD, et al. Antithrombotic Therapy after Acute Coronary Syndrome or PCI in Atrial Fibrillation (AUGUSTUS). N Engl J Med. 2019;380(16):1509–1524. https://doi.org/10.1056/NEJMoa1817083

19. Vranckx P, et al. Edoxaban-based versus vitamin K antagonist-based antithrombotic regimen after successful coronary stenting in patients with atrial fibrillation (ENTRUST-AF PCI). Lancet. 2019;394(10206):1335–1343. https://doi.org/10.1016/S0140-6736(19)31872-0

20. Yasuda S, et al. Antithrombotic Therapy for Atrial Fibrillation with Stable Coronary Disease (AFIRE). N Engl J Med. 2019;381(12):1103–1113. https://doi.org/10.1056/NEJMoa1904143

21. Lopes RD, et al. Optimal Antithrombotic Regimens for Patients With Atrial Fibrillation Undergoing Percutaneous Coronary Intervention: An Updated Network Meta-analysis. JAMA Cardiol. 2019;4(8):747–755. https://doi.org/10.1001/jamacardio.2019.1880

22. Stanifer JW, et al. Apixaban Versus Warfarin in Patients With Atrial Fibrillation and Advanced Chronic Kidney Disease. Circulation. 2020;141(17):1384–1392. https://doi.org/10.1161/CIRCULATIONAHA.119.044059

23. Pokorney SD, et al. Apixaban for Patients With Atrial Fibrillation on Hemodialysis (RENAL-AF). Circulation. 2022;146(23):1735–1745. https://doi.org/10.1161/CIRCULATIONAHA.121.054990

24. Reinecke H, et al. A Randomized Controlled Trial Comparing Apixaban With Phenprocoumon in Patients on Chronic Hemodialysis: AXADIA-AFNET 8. Circulation. 2023;147(4):296–309. https://doi.org/10.1161/CIRCULATIONAHA.122.062779

25. De Vriese AS, et al. Safety and Efficacy of Vitamin K Antagonists versus Rivaroxaban in Hemodialysis Patients with Atrial Fibrillation: Valkyrie. J Am Soc Nephrol. 2021;32(6):1474–1483. https://doi.org/10.1681/ASN.2020111566

26. Siontis KC, et al. Outcomes Associated With Apixaban Use in Patients With End-Stage Kidney Disease and Atrial Fibrillation in the United States. Circulation. 2018;138(15):1519–1529. https://doi.org/10.1161/CIRCULATIONAHA.118.035418

27. Ha JT, et al. Benefits and Harms of Oral Anticoagulant Therapy in Chronic Kidney Disease. Ann Intern Med. 2019;171(3):181–189. https://doi.org/10.7326/M19-0087

28. Eikelboom JW, et al. Rivaroxaban with or without Aspirin in Stable Cardiovascular Disease (COMPASS). N Engl J Med. 2017;377(14):1319–1330. https://doi.org/10.1056/NEJMoa1709118

29. Steinberg BA, et al. Off-Label Dosing of Non-Vitamin K Antagonist Oral Anticoagulants and Adverse Outcomes: The ORBIT-AF II Registry. J Am Coll Cardiol. 2016;68(24):2597–2604. https://doi.org/10.1016/j.jacc.2016.09.966

30. Seiffge DJ, et al. Ischemic Stroke despite Oral Anticoagulant Therapy in Patients with Atrial Fibrillation. Ann Neurol. 2020;87(5):677–687. https://doi.org/10.1002/ana.25700

31. Osmancik P, et al. 4-Year Outcomes After Left Atrial Appendage Closure Versus Nonwarfarin Oral Anticoagulation for Atrial Fibrillation (PRAGUE-17). J Am Coll Cardiol. 2022;79(1):1–14. https://doi.org/10.1016/j.jacc.2021.10.023

32. Wazni OM, et al. Left Atrial Appendage Closure after Ablation for Atrial Fibrillation (OPTION). N Engl J Med. 2025;392(13):1277–1287. https://doi.org/10.1056/NEJMoa2408308

33. Hijazi Z, et al. The novel biomarker-based ABC-stroke risk score for patients with atrial fibrillation. Eur Heart J. 2016;37(20):1582–1590. https://doi.org/10.1093/eurheartj/ehw054

34. Hijazi Z, et al. The novel biomarker-based ABC-bleeding risk score for patients with atrial fibrillation. Lancet. 2016;387(10035):2302–2311. https://doi.org/10.1016/S0140-6736(16)00741-8

35. Piccini JP, et al. Safety of the oral factor XIa inhibitor asundexian compared with apixaban in patients with atrial fibrillation (PACIFIC-AF). Lancet. 2022;399(10333):1383–1390. https://doi.org/10.1016/S0140-6736(22)00456-1

36. Piccini JP, et al. Asundexian versus Apixaban in Patients with Atrial Fibrillation (OCEANIC-AF). N Engl J Med. 2025;392(1):23–32. https://doi.org/10.1056/NEJMoa2407105

37. Ruff CT, et al. Abelacimab versus Rivaroxaban in Patients with Atrial Fibrillation (AZALEA-TIMI 71). N Engl J Med. 2025;392(4):361–371. https://doi.org/10.1056/NEJMoa2406674

38. Sharma M, Shoamanesh A, Connolly SJ, et al. Asundexian for Secondary Stroke Prevention (OCEANIC-STROKE). Late-breaking presentation, International Stroke Conference 2026. Full publication pending.

39. Verhamme P, et al. Abelacimab for Prevention of Venous Thromboembolism (ANT-005 TKA). N Engl J Med. 2021;385(7):609–617. https://doi.org/10.1056/NEJMoa2105872

40. Weitz JI, et al. Milvexian for the Prevention of Venous Thromboembolism (AXIOMATIC-TKR). N Engl.

Explore the Cardio Concepts Library

The actual story is messier, more interesting, and in some ways more instructive about how we reason from biology to the bedside.

The question itself is a good one. The central tension in anticoagulation has always been this: the same pathway that forms pathological clots also forms the clots that prevent you from bleeding to death when you cut yourself. Warfarin didn’t care about this distinction - it suppressed the whole system. NOACs improved things considerably by targeting specific factors (Xa or thrombin), reducing intracranial hemorrhage dramatically compared to warfarin, but they didn’t resolve the fundamental trade-off. You still can’t be fully anticoagulated without accepting some bleeding risk.

FXI sits interestingly in this picture. The observation driving the whole field was epidemiological first: people with congenital FXI deficiency have very low rates of VTE - but they don’t bleed spontaneously, unlike patients with hemophilia A or B. That asymmetry suggested FXI might be disproportionately important for pathological clotting rather than primary hemostasis. If true, blocking it could give you anticoagulant effect without the full bleeding penalty.

That hypothesis generated a wave of clinical development that’s now producing results - some striking, some sobering, and a few that directly contradict each other depending on which patient population you’re looking at.

The earliest proof-of-concept came from surgical VTE prevention. The ANT-005 trial with abelacimab - a monoclonal antibody that locks FXI in its inactive zymogen form before activation - showed VTE rates of 4–5% compared to 22% with enoxaparin, with no increase in major bleeding. AXIOMATIC-TKR with milvexian, an oral FXIa inhibitor, showed similar dose-dependent VTE reduction. These were small phase 2 trials, but the signal was clean enough to feel like genuine biology.

The logic held in the orthopedic surgery context because postoperative VTE is mechanistically a “contact activation” story - stasis, tissue factor exposure from the surgical field, FXI-driven amplification of thrombin generation in a setting where primary hemostasis has already been achieved by surgical repair. Blocking that downstream amplification step without touching the initial hemostatic response makes conceptual sense.

Which made it easy - maybe too easy - to assume the same would hold in atrial fibrillation.

PACIFIC-AF was a phase 2 safety trial: asundexian, a small-molecule FXIa inhibitor, tested against apixaban in 755 AF patients. The bleeding signal was encouraging - roughly a third of the bleeding rate with the 20 mg dose versus apixaban. No efficacy comparison was possible; the trial wasn’t designed for it. But the bleeding result was enough to justify the phase 3.

OCEANIC-AF enrolled nearly 15,000 patients. Primary efficacy endpoint: stroke or systemic embolism. The trial was stopped early by the DSMB - not for benefit, but because asundexian was clearly inferior to apixaban on stroke prevention. HR 3.79. Nearly four times the stroke rate.

Meanwhile, major bleeding was substantially lower - HR 0.32.

So the drug did what it was designed to do on bleeding, and failed completely on stroke prevention. That’s not a formulation issue or a dosing accident. Something biological is going on here.

The question worth sitting with is why FXIa inhibition works in VTE but apparently fails in cardioembolic stroke.

The most plausible interpretation is that LAA thrombus formation in AF is not primarily a FXI-dependent process. Atrial thrombus builds up through stasis, endothelial dysfunction, and a prothrombotic milieu in the left atrial appendage - a milieu that drives thrombin generation and fibrin deposition through pathways that don’t rely as heavily on FXI amplification as venous thrombosis does. Inhibiting FXIa may reduce overall coagulation activity, but not enough of the right activity to prevent LAA thrombus formation and downstream embolization.

There’s also the separate question of whether asundexian achieved sufficient depth of FXI pathway inhibition in this context. These two explanations aren’t mutually exclusive, and OCEANIC-AF alone can’t distinguish between them. That’s where the rest of the dataset becomes essential.

AZALEA-TIMI 71 ran essentially in parallel with OCEANIC-AF, but with abelacimab - and the bleeding results look completely different.

Abelacimab works upstream of asundexian. Rather than blocking FXIa after it’s already been activated, abelacimab locks FXI in the zymogen form, preventing activation from happening in the first place. In 1,287 AF patients at high bleeding risk who weren’t good candidates for standard NOACs, abelacimab 150 mg monthly SC reduced major and CRNM bleeding by more than 60% versus rivaroxaban. GI bleeding fell by roughly 89%. Zero ICH events in the abelacimab arm. The trial was stopped early because the DSMB felt the bleeding benefit was large enough that continuing the rivaroxaban arm was difficult to justify.

AZALEA was not powered for efficacy - we genuinely don’t know what abelacimab does to stroke rates in AF. That’s what LILAC-TIMI 76 is for, currently enrolling.

But the contrast with OCEANIC-AF raises a real mechanistic question: if locking the FXI zymogen before activation produces more complete upstream suppression than inhibiting FXIa downstream, you might eventually get both adequate antithrombotic effect and preserved hemostasis in AF. One mechanism failed. A different mechanism - working at the same axis but one step earlier - produced a safety signal that’s hard to dismiss. Whether that translates into efficacy in LILAC is the question nobody can answer yet.

OCEANIC-STROKE complicates things further - or clarifies them, depending on how you look at it.

Asundexian 50 mg daily, added to antiplatelet therapy in patients with recent non-cardioembolic ischemic stroke or high-risk TIA, reduced recurrent ischemic stroke by 26% (csHR ~0.74) with no increase in major bleeding (csHR 1.10, p=0.46). First phase 3 FXIa inhibitor trial to hit both endpoints simultaneously.

This is the same molecule that was stopped for futility in AF, now showing meaningful stroke prevention in a different population.

That pattern is worth thinking about carefully, because it’s not random. FXIa inhibition appears effective when the dominant thrombotic mechanism involves intrinsic pathway amplification at an atherosclerotic plaque - the kind of thrombus that builds on top of disrupted endothelium, where contact activation and FXI-driven thrombin amplification play a real role. It appears insufficient when the primary source of embolism is a preformed cardioembolic thrombus in the left atrium, where the underlying biology of thrombus formation runs through different amplification loops.

VTE, atherothrombotic stroke, and AF-related stroke share coagulation machinery - but which steps in that machinery are rate-limiting differs enough that the same target can be decisive in one setting and essentially irrelevant in another.

LIBREXIA-AF will answer whether OCEANIC-AF’s failure was specific to asundexian or a class-wide ceiling. Milvexian, another oral FXIa inhibitor, is being tested against apixaban in approximately 15,500 AF patients, with topline results expected 2026–2027.

If milvexian also shows inferior stroke prevention in AF, that’s reasonably strong evidence that FXIa inhibition as a mechanism - at least with small-molecule inhibitors - cannot deliver adequate cardioembolic stroke prevention, regardless of what it does for bleeding. At that point, the abelacimab story in LILAC-TIMI 76 becomes pivotal: does blocking FXI before activation achieve what inhibiting FXIa after activation couldn’t?

If LIBREXIA-AF somehow succeeds, the implications run in the other direction - asundexian’s failure becomes a molecule-specific or dose-specific story, not a class story. The field will have learned something different.

Either way, we’ll have a clearer picture of where the actual ceiling is.

One thing worth flagging before closing, because I think it gets underplayed.

OCEANIC-AF’s failure doesn’t mean FXI inhibition has no role in AF. It means asundexian alone is not sufficient to replace standard NOAC therapy for stroke prevention. Those are meaningfully different claims.

For AF patients who genuinely can’t tolerate standard NOACs - recurrent major GI bleeds, high-risk intracranial lesions, prior ICH - the AZALEA data suggests abelacimab could offer meaningful bleeding reduction while still providing some degree of anticoagulant coverage, accepting that the absolute stroke prevention may be less than with full-dose apixaban. That’s a real clinical trade-off, and for a specific subset of patients, it’s a legitimate one.

There’s a tendency in how we discuss trial failures to collapse everything: drug failed, class failed, idea failed. The actual picture is usually narrower. The question of which agent best prevents stroke in an average AF patient is different from the question of which agent is best for a patient who’s already bled on two different NOACs. Keeping those questions separate is what phenotype-based reasoning is supposed to do.

Subscribe now

So where does this leave us?

The FXI inhibitor story hasn’t delivered on its original promise - not yet, and maybe not in AF specifically. But it’s done something arguably more valuable: it’s forced a clearer articulation of why different thrombotic contexts require different therapeutic approaches. OCEANIC-AF failed, but the failure was informative in a way that a simple negative trial often isn’t. OCEANIC-STROKE succeeded in a different context with the same drug. That contrast, more than any single result, is what advances the field.

Mechanism matters. And so does knowing which mechanism you’re actually fighting - because the coagulation cascade is not one problem with one solution.

The story of FXI inhibitors is not simply about a new anticoagulant target. It’s a reminder that the same pathway can produce very different outcomes depending on the biology of the thrombus we’re trying to prevent.

I’ve summarized the key ideas in the framework below.

The Fxi Paradox

1.86MB ∙ PDF file

Download

Download

—Cardio Concepts

Previous When the Drug Is Right but the Risk Isn't Gone: Residual Thrombotic Risk in the NOAC Era

Next Looking Back: When Anticoagulation Stopped Being One-Size-Fits-All

References

1. Piccini JP, et al. Safety of the oral factor XIa inhibitor asundexian compared with apixaban in patients with atrial fibrillation (PACIFIC-AF). Lancet. 2022;399(10333):1383–1390. https://doi.org/10.1016/S0140-6736(22)00456-1

2. Piccini JP, et al. Asundexian versus Apixaban in Patients with Atrial Fibrillation (OCEANIC-AF). N Engl J Med. 2025;392(1):23–32. https://doi.org/10.1056/NEJMoa2407105

3. Ruff CT, et al. Abelacimab versus Rivaroxaban in Patients with Atrial Fibrillation (AZALEA-TIMI 71). N Engl J Med. 2025;392(4):361–371. https://doi.org/10.1056/NEJMoa2406674

4. Shoamanesh A, et al. Factor XIa inhibition with asundexian after acute non-cardioembolic ischaemic stroke (PACIFIC-Stroke). Lancet. 2022;400(10357):997–1007. https://doi.org/10.1016/S0140-6736(22)01588-4

5. Sharma M, et al. Asundexian for Secondary Stroke Prevention (OCEANIC-STROKE). Late-breaking presentation, International Stroke Conference 2026. Full publication pending.

6. Verhamme P, Weitz JI, et al. Abelacimab for Prevention of Venous Thromboembolism (ANT-005 TKA). N Engl J Med. 2021;385(7):609–617. https://doi.org/10.1056/NEJMoa2105872

7. Weitz JI, et al. Milvexian for the Prevention of Venous Thromboembolism (AXIOMATIC-TKR). N Engl J Med. 2021;385(23):2161–2172. https://doi.org/10.1056/NEJMoa2113194

8. Jain SS, et al. Milvexian vs apixaban for stroke prevention in atrial fibrillation: The LIBREXIA atrial fibrillation trial rationale and design. Am Heart J. 2024;277:145–158. https://doi.org/10.1016/j.ahj.2024.08.011

   Share

   Leave a comment

   Explore the Cardio Concepts Library

There’s a version of anticoagulation medicine that felt almost clean. You start a NOAC. You’ve addressed the main mechanism. You move on.

The problem is that stroke keeps happening.

Not often. Not in the majority. But in a meaningful subset of patients on appropriate anticoagulation - correct drug, correct dose, reasonably good adherence - ischemic events still occur. And once you’ve taken care of the obvious variable, you’re left with a harder question: what else is driving this risk?

This is roughly what “residual thrombotic risk” means in the AF context. It’s not a single mechanism and it doesn’t point to a single fix. But it’s probably the most underappreciated dimension of anticoagulation management right now.

The NOAC era didn’t eliminate stroke. It shifted where the remaining risk lives.

The pivotal trials - RE-LY, ROCKET-AF, ARISTOTLE, ENGAGE AF-TIMI 48 - established that NOACs reduce stroke risk significantly versus warfarin. The Ruff meta-analysis across 71,683 patients put the headline number at a 19% relative reduction in stroke and systemic embolism. That’s meaningful. But what remains is still a residual annual event rate that accumulates over years.

Some of that residual risk is pharmacological - inconsistent drug exposure, timing of doses, drug interactions. But some of it reflects something harder to fix: the coagulation cascade is not the only pathway generating thrombus in a fibrillating atrium. Blood stasis in the left atrial appendage, endothelial dysfunction, atrial cardiomyopathy changes that predispose to thrombus independent of rhythm - these don’t disappear because you’ve inhibited FXa.

COMPASS wasn’t about AF. But the question it asked was.

COMPASS enrolled 27,395 patients with stable atherosclerotic disease. Eikelboom and colleagues randomized patients to rivaroxaban 2.5 mg twice daily plus aspirin versus aspirin alone. The combination reduced the composite of stroke, MI, and cardiovascular death (HR 0.76), with a 42% relative reduction in stroke.

The point isn’t to import this into AF management - in the AF-PCI context, AUGUSTUS showed clearly that adding aspirin nearly doubles bleeding risk without meaningful ischemic benefit. The point is the underlying logic: atherothrombosis runs through at least two mechanistically distinct pathways, and suppressing one leaves the other active. Dual pathway inhibition worked because the residual risk after single-pathway therapy was real.

For AF, the question becomes parallel: after adequate anticoagulation, what’s still generating events, and is it something pharmacology can reach at all?

Biomarkers started pointing toward a more biological definition of risk

CHA₂DS₂-VASc was always a clinical approximation - age, hypertension, diabetes as imperfect proxies for the actual biology driving embolic events. The ABC-Stroke score, derived from ARISTOTLE and validated in RE-LY by Hijazi and colleagues, swapped some of those proxies for direct markers: NT-proBNP, high-sensitivity troponin T, age, and prior stroke.

Discrimination improved modestly (C-index 0.68 versus 0.62 for CHA₂DS₂-VASc), but the more interesting question is what the markers are actually measuring. NT-proBNP reflects atrial wall stress and cardiomyopathy progression. High-sensitivity troponin captures ongoing myocardial injury. These aren’t just better predictors - they’re pointing at a different disease process. The ABC-Bleeding score followed the same logic, adding GDF-15 (a marker of cellular stress and biological aging), and outperformed HAS-BLED.

Practically, what this suggests is that some patients carrying elevated biomarkers despite being on the right anticoagulant at the right dose aren’t failing anticoagulation - they’re failing because their atria keep remodeling, and the substrate for thrombus formation keeps regenerating. The drug isn’t wrong. The biology isn’t stopping.

Breakthrough stroke is where this becomes most concrete

Seiffge and colleagues pooled data from seven observational cohorts: 5,314 patients who had an ischemic stroke while already on OAC. Their annualized recurrence rate - 7.3% - was higher than what you’d expect from the broader anticoagulated AF population. And switching to a different OAC after the event didn’t reduce recurrence in any meaningful way (adjusted HR 1.2, non-significant).

Important caveats: this is retrospective, pooled, observational data with substantial confounding by indication. Patients who break through on OAC are not a random sample - they probably have more severe underlying atrial disease to begin with. So the recurrence rate can’t be read as causal, and neither can the null result from switching drugs.

But the clinical implication is still worth taking seriously: if you’ve already had a stroke on a properly dosed NOAC, the answer probably isn’t a different NOAC. Something else is driving it. And if you accept that framing, the logical next question is whether the source is structural.

LAA closure as a phenotype-specific approach

Most cardioembolic stroke in AF originates from the left atrial appendage. For patients at high thrombotic risk who can’t safely maintain long-term anticoagulation - or who have already failed it - addressing the anatomical source directly has a different appeal than escalating pharmacology.

PRAGUE-17 randomized 402 high-risk AF patients (CHA₂DS₂-VASc ≥3, or prior stroke or major bleeding) to LAA closure versus DOAC, with four-year follow-up. The composite outcome - stroke, TIA, systemic embolism, cardiovascular death, major bleeding - was non-inferior for LAAC (sHR 0.81). Long-term non-procedural bleeding was lower in the LAAC group (3.4% vs 5.9% per year, p=0.039).

OPTION brought this into the ablation context. In 1,600 patients undergoing catheter ablation, LAA closure at the time of the procedure versus continued OAC showed non-inferiority for efficacy and a meaningful safety advantage - non-procedural bleeding 8.5% versus 18.1%. For patients post-ablation who want to stop long-term anticoagulation, this is real data now, not just theoretical appeal.

Neither trial makes a case for replacing anticoagulation broadly. What they define is a more specific phenotype: high thrombotic risk, high bleeding risk simultaneously, where the LAA is likely the dominant source and where continuous systemic anticoagulation is either not tolerated or has already failed.

Some residual risk is iatrogenic

Not all of what gets called residual thrombotic risk comes from disease biology. Some of it comes from the prescription.

The ORBIT-AF II registry followed 5,738 AF patients on NOACs in routine US practice. About 9.4% were underdosed relative to FDA label criteria; 3.4% were overdosed. Underdosing increased cardiovascular mortality (adjusted HR 1.91) without reducing stroke rates. Overdosing increased major bleeding (adjusted HR 2.17) without improving efficacy.

Off-label underdosing often happens for reasons that feel protective - the patient is older, or has borderline renal function, or falls occasionally. The intent is reasonable. But the outcome is a patient exposed to some anticoagulant burden without adequate embolic protection. Residual thrombotic risk, created by the prescription rather than the disease.

This is probably more common than most clinicians would guess. And unlike atrial cardiomyopathy or LAA anatomy, it’s immediately correctable.

Subscribe now

Where this leaves the clinical conversation

Getting a patient on an appropriate NOAC isn’t the end of the risk conversation - it opens a different one. For the subset where that still isn’t enough, the relevant questions change: Are biomarkers signaling ongoing atrial remodeling that anticoagulation can’t address? Is the LAA the dominant anatomical source in someone who can’t tolerate long-term drugs? Is there a dosing gap that’s creating preventable risk?

The tools for approaching these questions are better than they were five years ago. The harder part is recognizing that the drug being right doesn’t mean the problem is solved - and being willing to ask what else is driving events before defaulting to switching medications.

One idea stayed with me while writing this piece:

Residual thrombotic risk is not one problem. Some risk comes from prescribing, some from biology, and some from anatomy.

I’ve summarized that concept into a one-page framework that you can save for quick review. PDF attached below.

When The Drug Is Right But The Risk Remains

287KB ∙ PDF file

Download

Download

—Cardio Concepts

Previous CKD and Anticoagulation: The Evidence We Wish We Had

Next The Cardiology of Wishful Thinking: What FXI Inhibitors Are Actually Teaching Us

References

1. Ruff CT, Giugliano RP, Braunwald E, et al. Comparison of the efficacy and safety of new oral anticoagulants with warfarin in patients with atrial fibrillation: a meta-analysis of randomised trials. Lancet. 2014;383(9921):955–962. https://doi.org/10.1016/S0140-6736(13)62343-0

2. Eikelboom JW, Connolly SJ, Bosch J, et al. Rivaroxaban with or without Aspirin in Stable Cardiovascular Disease. N Engl J Med. 2017;377(14):1319–1330. https://doi.org/10.1056/NEJMoa1709118

3. Lopes RD, Heizer G, Aronson R, et al. Antithrombotic Therapy after Acute Coronary Syndrome or PCI in Atrial Fibrillation (AUGUSTUS). N Engl J Med. 2019;380(16):1509–1524. https://doi.org/10.1056/NEJMoa1817083

4. Hijazi Z, Oldgren J, Lindback J, et al. The novel biomarker-based ABC (age, biomarkers, clinical history)-stroke risk score for patients with atrial fibrillation: a derivation and validation study. Eur Heart J. 2016;37(20):1582–1590. https://doi.org/10.1093/eurheartj/ehw054

5. Hijazi Z, Oldgren J, Lindbäck J, et al. The novel biomarker-based ABC (age, biomarkers, clinical history)-bleeding risk score for patients with atrial fibrillation: a derivation and validation study. Lancet. 2016;387(10035):2302–2311. https://doi.org/10.1016/S0140-6736(16)00741-8

6. Seiffge DJ, De Marchis GM, Koga M, et al. Ischemic Stroke despite Oral Anticoagulant Therapy in Patients with Atrial Fibrillation. Ann Neurol. 2020;87(5):677–687. https://doi.org/10.1002/ana.25700

7. Osmancik P, Herman D, Neuzil P, et al. 4-Year Outcomes After Left Atrial Appendage Closure Versus Nonwarfarin Oral Anticoagulation for Atrial Fibrillation. J Am Coll Cardiol. 2022;79(1):1–14. https://doi.org/10.1016/j.jacc.2021.10.023

8. Wazni OM, Saliba WI, Nair DG, et al. Left Atrial Appendage Closure after Ablation for Atrial Fibrillation. N Engl J Med. 2025;392(13):1277–1287. https://doi.org/10.1056/NEJMoa2408308

9. Steinberg BA, Shrader P, Thomas L, et al. Off-Label Dosing of Non-Vitamin K Antagonist Oral Anticoagulants and Adverse Outcomes: The ORBIT-AF II Registry. J Am Coll Cardiol. 2016;68(24):2597–2604. https://doi.org/10.1016/j.jacc.2016.09.966

   Share

   Leave a comment

   Explore the Cardio Concepts Library