“These are the patients that no one cares about. Well, apart from their parents.”

Developing a new drug is expensive, and large pharmaceutical companies cannot afford to spend hundreds of billions on developing treatments for ultra-rare diagnoses that affect only a few dozen patients worldwide. The revenues generated are simply too small. The system is broken, especially for conditions that hardly any clinician has even heard of. Yet, those are precisely the patients she wants to help. And it gets me thinking about fish.

Many years ago, someone asked me a question:

“I’m out on a boat. I cast a net and catch a number of fish. The net is the scientific method, and the fish are knowledge. But what are the fish that my net cannot catch?”1

The question stuck with me. We have to draw boundaries around concepts for them to be useful. Clear definitions facilitate communication. However, all important concepts are inherently complex, and therefore meaningful things will inevitably end up on the outside of the definition. Those are the ones we fail to catch.

The pharmaceutical industry has produced miraculous treatments: tyrosine kinase inhibitors for chronic myeloid leukaemia, proton pump inhibitors for peptic ulcers, antibiotics and much more. The list is as long as the years of lifespan humanity has gained in the past century. Humanity has built a machinery of progress for most of what ails us. However, none of that matters to the millions of people worldwide living with an ultra-rare condition. They swim past the net at the mouth of the machinery, unnoticed.

We also have nets that define what health care is responsible for. Many important fish are never caught, even though the health care net becomes larger and finer-meshed each year. The social determinants of health are the most significant influence on how long and how well we live, yet many of these determinants fall outside the scope of health care. There are good reasons for that. Different waters and different fish require different nets. Creating one net for all types of fish would be difficult and costly, which means that some important fish remain outside it. Similarly, it would be just as difficult and expensive to establish a health care system capable of encompassing education, housing, and culture, which is why certain important factors must remain outside of health care.

And we also have nets for what counts as good care. We measure healthcare-associated infections, bed days, accessibility, patient satisfaction, and many other metrics. However, the unique, unmeasurable aspects that matter most to an individual’s quality of life often slip through these nets. They don’t show up in our data, and we rarely have the time or resources to understand them. Clear boundaries make communication easier. The nets allow us to coordinate, capture, and measure, but sometimes they miss the point entirely.

I can’t fish. A friend once took me out on a fishing trip, and my expectations were high. I read up beforehand and pictured us coming home with a full box of fish. But, as mentioned, I can’t fish. After spending several hours by the water, we ended up eating something out of his freezer that evening. It was a quiet and lovely afternoon all the same.

Now, I should be careful about offering advice on fishing nets, having fished exactly once and achieved a hundred percent failure rate. I never did learn to fish. However, I did learn that precise definitions have their limits. Sometimes, the fish we cannot catch matter just as much as the ones that end up in the net.

This column was originally published in the Swedish Medical Journal, Läkartidningen, here

This is part 2 of a series of articles on the rare disease pharmaceutical industry. You may find it helpful to start with part 1.

I. The rare disease industry is big

Let’s assess the industry in three dimensions: patients, R&D, and healthcare spend / pharma revenue.

A. Patients

• 6000 - 8000 rare diseases exist, with ~95% lacking an approved treatment1

• ~3% (300M) of the global population have these rare diseases. This can be compared to cardiovascular disease (~6%), diabetes (~8%), chronic kidney disease (10%), and obesity (16%)2

• 80% of these cases are caused by 150 “more common” rare diseases (prevalence of 1-5 per 10 000)3

B. R&D

• 32% of assets in development target a rare disease. Currently, ~790 rare diseases are targeted (growing by ~7% per year).4

• ~44% of clinical trials in 2024 targeted rare conditions.5 This number seems high, and I haven’t been able to triangulate with other sources. It seems high despite being ~12% in 2006-20126, and growing rapidly thereafter.7 This aligns with the rapid growth of FDA ODD designations (from 150 in 2008 to 350 in 2019), but is sensitive to the definitions used.8

• ~50% of FDA NME/BLA approvals are for drugs with an orphan designation (see part 1). Regulatory agencies around the world report similar numbers.9 30-40% of orphan approvals target oncological indications.10

C. Financing

• ~7.5% of healthcare’s drug expenditure is on orphan drugs: 6.6% of EU in 202211 and 8% of US in 2016 and 202212

• Rare disease drugs make up a ~$200B market; ~13% of pharma’s total $1.5T13

• 20-26% of pharma’s prescription drug revenue comes from orphan drugs14

In brief: despite afflicting relatively few patients, rare diseases make up an oversized share of the global biotech pipeline, clinical trials and healthcare spend. This is why a large share of venture capital goes into rare therapies.15

II. Bayesian priors for rare disease drugs

Bayesian priors are useful for clinicians, investors, and anyone dealing with uncertainty. The past doesn’t predict the future, especially in biotech, but it can help calibrate intuition and improve probability estimates of future events.

These are the best estimates I have been able to find for rare disease assets. Interpret cautiously and apply judiciously, as every data point is limited by the original data (as well as definitions, time frames and methodological limits).

A. Clinical trials for rare diseases are…

• Significantly smaller trials (median no of patients 29 vs 6216, 65 vs 15017, 187 vs 52118 ; depending on timing and methodology) with increasingly smaller sizes for later stage trials19

• More often single-arm (63% vs 30%)20

• More often non-randomized (65% vs 36%)21

• More often open label (79% vs 52%)22

• More difficult in phase II and III23

  • 23% longer start-up timelines

  • 30% more planned visits

  • 19% longer treatment durations

  • 30% lower completion rates

B. Regulatory

• Rare drugs have a 2x higher approval probability than an average phase 1 asset (17% vs 8%).24 The difference seems mainly driven by better probabilities in phase 1 and phase 225, probably due to better understanding of genetic causes (a genetically supported target doubles clinical development success26). I haven’t seen the breakdown, but I assume that the large share of oncology (has a lower LoA of ~5%) within rare in part counteracts the genetic improvement.

  

• Orphan drugs need fewer trials to get approved (1.5 compared to non-orphans’ 2.4).27 71% of orphan assets in EU were approved by EMA on a single trial28

Being aware of both A and B above is important. Single-arm approvals and genetic basis raise the likelihood of approval. However, they also increase some of the potential failure risk to post-approval (confirmatory trials and accelerated-approval withdrawals). This contributes to the late acquisition profile (more in part D below).

C. ommercially, rare disease drugs…

• Quicker peak sales: 4 years vs 10 years29

• Have less competition: ~40% of rare approvals in 2022-2024 were first-in-disease30

• Make up ~25% of all blockbusters31

D. is for dealmaking

Orphan drugs…

• Have less costly drug development: 23% of non-orphan R&D spend (note: average number, highly dependent on modality). Phase III cost of an orphan drug is often estimated to be ~25% of that of a non-orphan drug.32

• Have same time until exit (median 8 yrs), capital invested until exit ($50-60m), multiple on invested capital (~5.5x)33

• Are acquired later (65% in phase 3/approval vs 31% for non-orphan)34.

E. Public market view on ODD

• Valued by markets: For US-listed biotechs, ~3% increase in share price (larger effect in nano (<$50m) and micro-sized ($50-250m) biotech35

• Higher ROA: Publicly listed companies with orphan drug market authorization have ~10% higher ROA than non-orphan drug companies; and ~5x higher operating profit36

• Often commercially underestimated by analysts: 67% of launches met or exceeded analyst expectations in launch year37

III. How to use this

These facts are not prescriptive. They do not inform investors on whether to invest more or less in rare diseases, or conclude whether the current regulatory frameworks are effective. The data points are inherently limited: they are historical, they are averages, and any new treatment being developed will per definition not be in the dataset.

However, they can be highly useful, especially for investors early in their career, when assessing new potential assets targeting rare diseases, as prior probabilities to help calibrate intuition. In other words, before adjusting for the specific details of an individual investment case, one can estimate that a rare disease asset, compared to non-rare assets, has a higher probability of reaching approval, but will most likely need to be taken further before partnering with a larger pharma.

Also, keep in mind that all priors are averages and that the underlying data is often a wide range. Priors, like the increased likelihood of reaching approval, should be adjusted for other factors which independently affect the outcome, for instance therapeutic area, modality, and the strength and extent of supporting data.

I hope you find this useful and that it helps investors make better decisions, so more treatments can reach patients. If you have any useful datapoints or data sets, please share them and I’ll add them to this article.

Thanks to Torquil Jackson, Eleonora Lugarà and Jessica Vitos for reading drafts and providing useful feedback.

How did medicine, healthcare and pharma end up here?

Today’s situation was far from inevitable. To understand the rare & orphan drug industry, we must begin from the beginning, in 1957, with a story of sedatives, rookies and Nazis.

I. Sedatives, Rookies & Nazis (1957-1983)

In 1957, Chemie Grünenthal launched a new sedative, Distaval, treating anxiety, poor sleep and morning sickness. Its marketed safety contributed to its rapid success. By 1960, ~15 tons were sold annually1 , making it the second best-selling drug in Germany, surpassed only by Bayer’s Aspirin.2

The drug spread rapidly across Europe and globally, and the next step was approval in the US. In 1960, an American licensee sent the FDA an application, which came to Dr. Frances Kelsey, a 46-year-old GP with a PhD in pharmacology. She had reviewed scientific articles for years3, but had joined the FDA only one month earlier and had never reviewed a drug before. She was a rookie.

At this time, there were limited requirements on what data had to be submitted for new drugs (for instance, manufacturers did not need to prove efficacy). However, the team assessing the application (Kelsey, a pharmacologist and a chemist) were hesitant about the limited absorption and excretion data and the short chronic toxicity studies. By today’s standards the data was very limited, and insufficient for concluding that chronic use was safe4.

Kelsey found no data proving the drug was unsafe, but she was still hesitant. She had previously worked with Professor Eugene Geiling on investigating unusual deaths caused by an improperly formulated antibiotic5, which likely affected her gut feeling. At this point, Kelsey didn’t want to approve the application, but she lacked the evidence to reject it. So she asked Barbara Moulton for advice. Barbara was a “chain-smoking straight-shooter” who recently whistle-blew about corruption, caving to industry pressure and lax standards at the FDA.6

Barbara told Kelsey about a loophole: a third option besides approving or rejecting the application: ask for more data. The FDA had 60 days to review applications, but if more data was requested, the submitting company had to comply, and the clock reset to 60 days.

And that’s exactly what Kelsey did. The company got frustrated and complained to her superiors. But Kelsey stood her ground despite the drug already being approved in several other countries. In 1960, Kelsey read a BMJ case series describing four patients with peripheral neuropathy (a type of nerve damage) from the drug. She continued standing her ground, in total, asking for more data four times.7

Her doubts were confirmed in 1961. An Australian obstetrician published a paper linking the drug with congenital (birth) defects. The news spread quickly. The manufacturer objected ardently, resisting it in every way possible8, but the drug, thalidomide, was finally withdrawn from the market that year.

Due to the widespread use of thalidomide, more than 10 0000 infants in 46 countries are estimated to have been born with thalidomide-related defects, of which up to 40% died within one year of age.9 Unnecessary, terrible, and preventable suffering.

At that time, the US company had already distributed over 2.5 million tablets to over 1 000 US doctors on an “investigational basis” (i.e. for an ongoing study). The doctors, in turn, gave thalidomide to nearly 20 000 patients, including hundreds of pregnant women. Many more would have been hurt if not for one bold scientist standing her ground.

I always thought that the thalidomide tragedy was a story about the challenges of biology and detecting rare side effects.

However, it is not. It is a story of limited regulatory demands, limited scientific analysis, purposeful deception and Nazis. Grünenthal tried to cover up side effect signals like peripheral neuropathy:

“They denied them, they minimised them, they pressured editors of medical journals not to publish critical articles, they hired private detectives to spy on doctors…”.10

Some data suggests thalidomide was initially tested on concentration camp prisoners as an antidote to nerve gas, prior to being developed as a sedative.11 Adding insult to injury, Chemie Grünenthal did not apologize or accept blame for the thalidomide tragedy until 2012, when they founded the “Grünenthal Foundation for the Support of Thalidomide-affected People”.12

Finally, several prominent Nazis held leadership positions in the company before, during and after this period. It seems likely that they held a different view on basic tenets of medicine, like the universal value of human life.

• Heinrich Mückter, Head of Research at the time, had previously led experiments on concentration camp prisoners at Buchenwald.13

• Martin Staemmler, Head of Pathology at the time, was a strong Nazi supporter and honorary professor in racial hygiene. After Germany’s invasion of Poland, he worked with the SS on population policy, deciding who should live and who should die.14

• Otto Ambros, a Nazi convicted of mass murder at the Nuremberg trials who was later freed, became chairman of Grünenthal’s advisory committee. His previous accomplishments included co-inventing sarin gas for the Nazis.15

The story of Frances Kelsey and thalidomide is worth reading for many reasons. In this specific case, it sets the backdrop for the rare disease industry.

Kelsey’s refusal to approve thalidomide led to the Kefauver-Harris Amendment (the “Drug Efficacy Amendment”), signed into law in 1962 (unfortunately immortalizing the two congressmen’s names and omitting Kelsey). Now drug manufacturers had to provide more proof of their drugs’ effectiveness and safety before approval. Previously, investigational-use drugs could be distributed freely if labeled as such. Now, all drug sponsors had to submit an IND (investigational new drug) application before starting a clinical trial.16

This may sound obvious today, but it was a tremendous change at the time. Kefauver-Harris raised the bar for pharmaceutical companies, making drug development more expensive, and many previously marketed drugs were abandoned (black segment in figure below).

After 1962, drugs for a rare disease affecting a small number of patients were evaluated and priced just like drugs targeting the most common conditions. It made little sense for pharmaceutical companies to develop drugs for rare conditions.17

II. How Sausages & Orphan Industries Are Made (1983-2000)

Few drugs had been developed for rare diseases, and even fewer were developed after Kefauver-Harris. The pharmaceutical industry focused on potential blockbusters for common conditions like hypertension. By 1979, the FDA knew the system was broken: drugs weren’t being developed for rare conditions. Many patients are afflicted by rare diseases, as there are many rare diseases, even if each individual rare disease is rare. Patients and patient organizations were frustrated. US activist organizations like NORD lobbied politicians and lawmakers, who were investigating the problem, holding hearings and drafting legislation.18

Fortuitously, an L.A. Times reporter listened to a hearing and wrote a short article. Jack Klugman, a TV celebrity and star on the medical drama Quincy M.E, happened to read it.

His brother Maurice had a rare cancer and they both joined the movement,19 contributing to two episodes of his show on rare diseases.20 This is remembered by many as the moment when public support grew for solving the issue of orphan drugs: rare conditions for which no drug is available.21

In a perfect storm of bipartisanship and lobbying (including celebrity testimony), in 1983 the ODA (Orphan Drug Act) was passed. President Reagan, a strong proponent of deregulation, signed it enthusiastically22, seeing it as a catalyst for innovation, not a barrier. For a condition affecting fewer than 200 000 people23, the act gave manufacturers three main benefits:

(1) Subsidies for clinical trials;

(2) Tax credits of 50% for clinical development costs (25% since 2017); and

(3) Exclusivity to market the orphan drug for 7 years post-approval

(1) and (2) eased capital requirements, and (3) increased potential profits for an approved drug. It was suddenly possible (albeit still uncertain) for manufacturers to recoup the costs of developing orphan drugs.

-

This is the commonly told story. However, there is another version, penned by FDA insiders, which paints a slightly different picture.

After Kefauver-Harris, drug development costs and timelines skyrocketed, and the FDA faced a legitimacy crisis. However, there was limited awareness of and interest in pharmaceutical companies abandoning unprofitable drugs.24 The main critique was that the FDA was causing a drug lag compared to Europe, with a bureaucracy literally killing people by keeping life-saving medicines off the market. The public viewed the FDA as a cold, scientific gatekeeper. This was a problem for the agency and its stakeholders.

The orphan drug problem offered the FDA a unique opportunity to signal its care for patients. Beyond the financial incentives for drug manufacturers, the ODA created new political alliances. Suddenly, rare disease activists like Abbey Meyers, founder of NORD, became a fierce defender of the FDA.

(For those interested: read this review of Daniel Carpenter’s magnum opus on the FDA: Reputation and Power. The main lesson for our story is to remember the political dimension to the FDA’s actions, which is often opaque at the time and can only be understood in retrospect.)

Both stories converge at one point: the ODA changed everything. The industry responded to the incentives and the number of rare disease drugs developed and approved increased significantly.25

The percentage of orphan drug approvals increased from ~10% in 1984 to ~20% during 1990-2005, reaching ~33% by 2014 and ~50% in 2024.26

III. A Global Industry Is Born (2000-2020)

A. Orphan drug regulation goes global

Several countries followed the US. Japan’s orphan drug laws were enacted in 1993, Australia in 1997, and South Korea in 1998. In 2000, the EU, the second largest market after the US, followed. The EU Orphan Regulation created an equivalent system to the US, with some tweaks:

During this era, orphan drugs became the core of multi-billion dollar drug companies (e.g. Genzyme, BioMarin, Vertex, Alexion). Later, big pharma acquired these companies (Sanofi acquired Genzyme in 2011 for $20B, Takeda bought Shire in 2019 for $62B, AstraZeneca acquired Alexion for $39B in 2021). Orphan drugs had gone from being commercially unfeasible to an important strategy for big pharmaceutical companies.

China lacks an ODA equivalent, but rare disease support is spread across several laws and policies. In contrast to other geographies, China stipulates a catalog of conditions to define a rare disease, rather than a prevalence cut-off.27 For such conditions, foreign clinical data (waiving China-specific trials if ethnic representation is addressed) and real world evidence (if trials are unfeasible) can be accepted. These orphan drugs can qualify for regulatory pathway accelerations, such as priority review, breakthrough designation and reduced review timelines.

In general, the support structures for orphan drugs are less established in China than those in Europe and the US. This is reflected in a recent survey by TD Cowen of 45 BD executives, in which rare disease assets were highlighted as an area in which China is weaker.28

B. FDA launches additional programs

Since the ODA, the FDA has launched other programs to shorten the time for a drug targeting a serious condition to reach patients. These programs don’t necessarily target orphan drugs and can be used for non-orphan indications but quickly became very relevant for many orphan drugs. The table below summarizes the programs and their respective eligibility criteria and benefits.

Orphan drug candidates often have several of these designations. ~50% of approved orphan drugs have a breakthrough therapy designation, and ~90% of orphan drugs with a breakthrough therapy also get a priority review.29

This is another way to view the impact of the programs: how they impact the development timeline.

The PRV Program is worth highlighting. In 2012, priority review was extended with the PRV (Rare Pediatric Disease Priority Review Voucher) program.30 Companies developing a treatment for a rare pediatric disease can apply for a Rare Pediatric Disease Designation. If they qualify, they will earn a transferable voucher allowing priority review (6-month vs 10-month) upon approval. These vouchers are sold to big pharma companies (where those 4 months is worth a lot), with a recent market price of ~$150M.

The PRV seems to have had an impact. Between 2012 and 2024, 53 PRVs were awarded for therapies across 39 diseases, of which 36 did not have any FDA-approved treatment.31

The EU has additional mechanisms beyond the Orphan Regulation:

• 2007 EU Paediatric Regulation: Orphan drugs can get 6 additional months of patent protection and 2 more years of orphan exclusivity if the company performs the planned pediatric trials

• 2016 PRIME (Priority Medicine) scheme: early and dedicated EMA regulatory, scientific support, and accelerated assessment possibility (56% of PRIME drugs have an orphan designation)

These programs do not mean that developing a drug for a rare condition is now easy. Many challenges remain, e.g. difficulties diagnosing and recruiting patients to trials, limited data on natural history, large interpatient variation, uncertainty regarding payors’ willingness to pay and much more. But these programs have made it easier.

IV. An Evolving Industry (2020 - …)

Today, orphan drugs are a significant part of the pharma industry, making up ~50% of FDA approvals and generating ~$200B in annual revenue32

This chapter is being written today. It’s difficult to make predictions, especially about the future, but I believe these trends will form part of this chapter as it is being written:

1. More diagnoses & patients

~300M patients globally suffer from 6000 - 8000 rare diseases.33 The theoretical number of 300M is constrained by the costs to find a diagnosis, but the cost is decreasing.

A central diagnostic technology is genomic sequencing, analyzing DNA to find mutations or variations causing a disease.34 Between ~2008 and ~2012, growing adoption of massively parallel sequencing (MPS), also known as next generation sequencing, for analyzing DNA reduced the sequencing cost thousandfold.

After ~2015, the cost stabilized for years. Recently, new technologies (e.g. Ultima Genomics) have reduced the sequencing cost, with one company offering whole genome sequencing reagents for $80.35

This cost reduction, combined with sophisticated health care systems and increased health care expenditure, will lead to more patients being diagnosed. The speed of this change will differ across health systems. In Sweden, for instance, whole-genome sequencing is already performed on all cases of pediatric cancer and is frequently (and increasingly) used in the case of a suspected rare disease.36

2. CGT treatments for the untreatable

In the past decade, cell and gene therapies (CGT) have cured previously incurable diseases. A couple of examples:

1. Zolgensma was approved for spinal muscular atrophy (SMA) in 2019. SMA is a rare genetic disease that causes progressive muscle weakness and wasting, with many patients dying in their teenage years. Zolgensma is a one-time gene therapy, which causes motor neurons to express the missing SMN gene. It was the first gene therapy approved for SMA and has shown transformative clinical results. Despite not being curative, it was a guiding example of how advanced treatments for rare diseases could work biologically and commercially.

2. In 2023, Casgevy was approved for sickle cell disease. Sickle cell disease is a genetic blood condition where the red blood cells become hard and crescent-shaped and become stuck in vessels causing pain (vaso-occlusive crises). Until Casgevy, the only treatments were preventive (hydroxyurea, blood transfusions) and symptomatic, and the only possible cure was a hematopoeitic stem cell transplantation (with the risks that entails). Casgevy is a one-time CRISPR-edited gene therapy where you remove the patient’s stem cells, introduce a gene that creates functional fetal hemoglobin, and return the modified stem cells, preventing >97% of vaso-occlusive crises. Again, outstanding.

There are many examples (Lenmeldy for early onset MLD, Hemgenix for hemophilia B, Elevidys for Duchenne muscular dystrophy etc.), but these suffice to illustrate the paradigm shift: more previously untreatable conditions are now treatable.

3. Payor pushback on premium pricing

Many novel treatments for rare diseases are administered as one-time treatments. While producing tremendous impact, they have also had tremendous price tags of $2-4M per treatment.

One-time cures benefit patients, but the price tag challenges payors who must find funding and assess the benefit and health economic savings.37 Despite serving a small number of patients, orphan drugs already account for 6.6% of EU38 and 7.5-8%39 of US pharmaceutical spend.

The prices are high, but there’s no other way for manufacturers and investors to recoup the capital invested. Without those prices, the drugs wouldn’t get developed.40 This is reflected in the fact that orphan drug prices are inversely correlated to the number of patients.

Payors face the same challenge as when investing in preventive care: Where will the money come from? Does this additional cost have the highest political priority compared to all other spending options?

Pricing is a problem not just for payors, but also for biotechs. Bluebird bio’s Zynteglo can (virtually) cure beta-thalassemia. The treatment was priced at ~$1.8M in Europe in 2020 but the company failed to secure reimbursement, especially in Germany where payors pushed back on price. By 2022 the company withdrew the therapy and shut down its European operations after being unable to secure an acceptable price.

As launch prices rise, and more one-off curative treatments get approved, this will become an increasingly large challenge for payors.

4. Regulatory resolve & revisions

Regulatory changes are a direct consequence of the previous trends: a growing number of patients and a growing number of expensive treatments. This affects how regulators view the legislation. However, we have yet to see significant regulatory pushback; this has so far primarily happened at the payor level.

Continued support in the US:

• The PRV program got reauthorized (until 2029), with strong bipartisan support

• The FDA launched Rare Disease Evidence Principles (RDEP) in 2025 to facilitate development of treatments for ultra-rare diseases41, and published draft guidance on innovative study designs for CGT for small populations and post-approval studies for CGT

Revisions planned in the EU:

• The EU Pharma Package (provisionally approved in Dec ’25, final approval expected in 2026) will42

  • Reduce baseline orphan exclusivity from 10 to 9 years. New “breakthrough” orphans, which treat conditions with no previously existing treatment, will have a baseline of 11, with a possibility to extend to up to 13.

  • Prevent multiple orphan designations for the same active substance from generating separate exclusivity periods (to prevent evergreening of the same active ingredient), with a maximum of +2 years of orphan market exclusivity due to new indication.43

  • Remove the extra two-year market exclusivity for completing a paediatric investigation plan

The general regulatory direction is reflected in forecasts. Analysts believe that the orphan drug industry will continue to grow significantly, both in absolute terms and as a share of all prescribed drugs.

If you made it this far, you found this topic as fascinating as I do. The history of orphan drugs shows us why patients with rare diseases should have more hope now than ever before, despite potential shifts in EU regulation.

Looking forward, the industry seems relatively robust. In the near future, it seems likely that rare drugs will continue expanding as a proportion of FDA approvals, healthcare budgets and pharmaceutical companies’ revenues. As of yet, there is no strong counterbalancing mechanism. It’s difficult for politicians to overtly argue against supporting drug development for rare diseases (as with cancer or any other serious disease). However, we have seen payor reluctance to reimburse expensive cures for rare diseases. That has upstream effects (e.g. uncertainty among investors and biotech entrepreneurs on commercial potential), but so far that hasn’t significantly counteracted the current regulatory incentives.

The automaton known as the rare disease industry is a thing of wonder and terrible beauty. Rare diseases form a unique therapeutic area, and will likely remain so for the near future.

V. Emmetropes

• The history (and the future) of the pharma industry is non-linear and malleable. It has and will continue being shaped by the political forces within and outside of the FDA. Be aware of the (for outsiders opaque) political game which can swing any way

• No orphan industry without the ODA, no ODA without the Kefauver-Harris, and no Kefauver-Harris without the thalidomide tragedy and patient champions like Frances Kelsey

• Patient organizations have changed the history of medicine and healthcare and should not be underestimated

• Rare diseases have several fundamentally different principles compared to other therapeutic areas: e.g. commercial & payor risks, clinical development, market access and protectability. These principles will remain different, but as regulation is updated, the consequences of these differences will change

This is part 1 of a series of articles on rare disease pharma. Click here to continue to part 2: The Unique Profile of Rare Disease Assets.

Thanks to Torquil Jackson, Eleonora Lugarà and Jessica Vitos for reading drafts and providing useful feedback.

8

Grünenthal insisted they were blameless: “The North Rhine-Westphalia public prosecutors found the company obstructive. They had to seize the most important Grünenthal documents in police raids on its “bunker” and a company lawyer’s house … Grünenthal came defiantly to court. Its battery of 40 lawyers, easily outnumbering the prosecution’s, did their best to delay. They several times threatened to walk out. Eighteen of them spent 12 days failing to shake the key testimony of the Hamburg paediatrician Dr Widukind Lenz, who had alerted Grünenthal to the tragedy.” Source

Sometimes data can be interpreted in different ways. Not in this case. The timing of the epidemics of malformations provided a natural experiment, and exceptionally strong support for a causal connection, which any reasonable scientist would accept as damning.

An excerpt from Lenz 1988:

“In Canada, thalidomide was put on the market in April 1961, and the dysmelia epidemics did not start before December 1961 (Webb, ‘63). The end of the epidemic likewise depended on the time the drug was withdrawn. There was, as expected, a sharp fall of the German epidemic in July and August, roughly 8-9 months after the withdrawal of thalidomide by the end of November 1961. In Brazil, thalidomide continued to be widely sold up to June 1962, and the epidemic ended about half a year later than in Germany.”

28

TD Cowen - Biopharma BD: Assessing China’s Innovation & Industry’s Evolving Appetite

41

There is no universal definition, but NICE defines it as ≤1: 50 000 births in England. No consensus definition exists for the US. According to an Orphanet (a reference database for rare diseases) study, 85% of rare diseases have a prevalence of < 1: 1 000 000. In other words, there are many rare diseases that truly are very, very rare.

Who can see the barely perceptible line between the man who can not read at all and the man who does not read at all? The literate who can, but does not, read, and the illiterate who neither does nor can?

- Joseph D. Eggleston Jr, 1910

If you work with scientific information, you will make better decisions with five scientifically trained colleagues available 24/7 to assess all available literature, extract relevant information, and synthesize it into a nuanced representation of our current knowledge.

If you, like me, can’t afford to employ five colleagues available 24/7, you should be using AI tools. The analogy is flawed, but captures my experience of using such tools.

In recent years, I have tried nearly every AI tool I have encountered for processing literature. These are the tools I have, so far, found most useful. However, first, brief context. Understanding science is tricky. Whenever you study a new scientific domain, you will initially be constrained in terms of:

• Domain expertise: you will not be an expert. Even with the aid of other domain experts, you will (and should) know less than the researchers in that field

• Time: you are always constrained by time (high opportunity cost to spending several years solely exploring one specific domain)

• Context: unless you happen to have previously worked in that field, you will not have a deeper understanding of the tacit nuances of a field

Finding ways to better understand science is useful. There are many ways to do this, for instance developing a deeper understanding of scientific methodology, metascience, Bayesian probability, epistemology, clinical trials or problem-solving methodologies. However, AI tools complement that: they automate the manual time spent on repetitive tasks like searching for articles, extracting and compiling data, and contextualizing findings.

AI tools save time. I wrote my first literature review ~15 years ago. I spent months manually finding and reviewing articles in PubMed, collecting and excluding them, and compiling everything in an Excel sheet. Using today’s AI tools, the same work would take ~2 weeks. The time spent reading and assessing data wouldn’t change, but much of the surrounding non-analytical work would be automated. The time gained can instead be spent on developing a deeper understanding and making better decisions.

To summarize the list below, I can recommend (no conflicts of interest):

1. Elicit: broad and deep questions

2. Consensus: specific and targeted questions

3. MS Co-pilot: ideation, discussion and devil’s advocate

I. Elicit

What: A powerful, flexible tool for extracting & compiling data, and summarizing articles

Pricing: Free (basic), $42-49/month (Pro)

Typical use case: You ask a prompt. The system offers advice to improve your question. You can select the source (research papers, clinical trials) and format (general review, 1-page overview, research gap analysis). You choose between fast, balanced or comprehensive search (i.e. more sources and time). Depending on your choices you’ll get different reports. If you choose systematic review you can fine-tune inclusion/exclusion and data extraction columns (and definitions) at each step.

Alternatively, you can choose a custom workflow (research landscape, clinical trial analysis, competitive landscape, or a general research agent). In this case, you can chat with the system and refine or complement the query in a chatbot interface.

Why it’s useful: Elicit is the most comprehensive in aggregation, nuanced in interpretation, and flexible for the right depth compared to most tools. Its main limitation is that of many other systems: data will often be extracted from abstracts (lacking details) if you don’t have institutional access.

Very useful for getting an initial understanding before diving into the literature.

II. Consensus.app

What: A focused tool for answering questions and summarizing the literature’s view

Pricing: Free (basic), $10/month (Pro), $44/month (Deep plan)

Typical use case: You have a focused question and enter it. Consensus will quickly write a summary answer (a “thread”), with useful summaries of the underlying articles. For most questions, it will also provide a consensus-meter (a colored bar chart showing if the literature agrees/disagrees with the answer). This may sound minor, but it can be useful in identifying and understanding areas of uncertainty or contention.

Why it’s useful: Sometimes you need a quick answer to a specific question, pathway, or hypothesis, and then Consensus is perfect. Its combination of a quick high-level view of the literature and clear representation of uncertainty makes it work well for specific narrow questions. Highly recommended.

III. Co-pilot

What: An AI service that Microsoft has forced on humanity using a distributional embuc, and which I never thought I would use as much as I do today.

Pricing: Free, $20-30/month (Pro)

Typical use case: You use a detailed prompt to get Co-pilot to reason as a specific role (e.g. drug developer, biostatistician, clin ops expert). Then, enter specific questions or a hypothesis and ask it to challenge your thinking and identify gaps.

Why it’s useful: The latest “deeper thinking” (GPT 5.4) offers surprisingly nuanced and well-grounded responses if you spend time finding suitable prompts. However, due to Co-pilot’s chatbot structure, it interprets the literature more freely than the other tools. This leads to less grounded assessments and conclusions (estimates or opinions) than the other tools, and I would therefore definitely not only use this (or any other generic chatbot) to gain a deeper understanding of a new scientific area.

IV. Other tools

Important: If you’re new to these tools, do not trust my recommendations.

First, in times of substantial change, you need to shift from the baseline exploit (do what’s worked well for you so far) to explore (experiment more than what you normally would do). So you need to be trying many tools.

Second, which tools are most suitable will depend on what work you do, and how you do that work. Before settling on the tools above, try many others. Some examples of tools scite.ai, amass.tech, researchpal.co, scholarcy.com and citation mapping systems: litmaps.com, researchrabbit.ai, connectedpapers.com.

Finally, a person who does not read is like one who cannot read. Something is definitely lost when AI tools guide one through the literature. However, time is gained, and if used judiciously, I believe the tools offer a clear net benefit.

V. Emmetropes

• Using AI tools resembles having five research assistants available 24/7. They cannot and should not replace all the things you do, but they can free up a lot of time

• These tools are valuable: the time they save can enable a deeper understanding

• Shift from exploit (work like you normally do) to explore (test new working methods) when technology or your field is rapidly changing

This is probably for the best. A realistic medical drama would be unwatchable. Few viewers would tune in eagerly to observe the administrative choreography of faxing a medical record from one provider to another. The Good Doctor, too, takes considerable liberties: it is heightened, overcomplicated, and generously infused with romance. But there is something special about it.

The series tells the story of Shaun Murphy, a surgical resident with autism and savant traits. He retains everything he has read, discerns patterns invisible to others, and arrives at diagnoses that bypass ordinary human cognition. Social interactions, however, are challenging. He struggles to read emotions and relationships. Fortunately, his colleagues step in, providing the interpersonal judgment and communication skills he lacks.

The idea is alluring: a mind that contains the totality of medical knowledge. Each year, countless patients are harmed by avoidable medical errors, mistakes that shouldn’t happen. There are also rare and perplexing cases when it can take weeks, months, or even years to separate signal from noise. In cases like these, both patients and colleagues would welcome a Shaun at their side.

However, the premise collapses under scrutiny. No human can carry the full breadth of medical knowledge in their head. Dr Shaun is entirely implausible. PubMed now indexes well over a million new scientific articles every year.1 The human brain did not evolve to manage immense data sets; natural selection had other priorities in mind.

Many years ago, I met an internist who resembled Shaun. He did not share Shaun’s social challenges, but rather his depth and extent of knowledge. He was an experienced and disarmingly modest physician who had recently returned to Sweden after working in the U.S.

In the breast pocket of his scrub top, he carried a Palm Pilot loaded with guidelines and articles. He knew an astonishing amount, and if he was uncertain, he quickly looked it up. The Palm Pilot is a relic, and has been so for many years, yet I find myself thinking of that doctor whenever I read about artificial intelligence, in the form of large language models, entering clinical medicine.

AI, of course, has serious limitations. It can hallucinate (an elegant euphemism for making things up), misread context, and present incorrect conclusions with great confidence. And yet, study after study suggests that AI systems can outperform physicians on measures of medical knowledge, make more accurate diagnoses in complex or rare cases2, and even provide responses to patient questions that are judged to be more empathetic and precise3.

But studies are like margaritas: best enjoyed with a pinch of salt.

In The Good Doctor, Shaun is not just a character, but also a symbol embodying the idea that knowledge alone is insufficient. Good medicine is not merely the recitation of collection of facts; it is a practice shaped by relationships, priorities, judgment, and a constant negotiation with uncertainty. The clinical encounter is more than diagnosis. Still, diagnosis matters greatly, especially in the cases that go wrong.

Many AI systems are trained to be honest, helpful, and harmless, not unlike Shaun himself. They become a kind of colleague endowed with near‑limitless knowledge - but one who generally cannot, and should not, bear responsibility for the care of a patient. But alongside human colleagues, us ordinary mortals who now and then forget things, they may nonetheless help make care better and safer.

The futurist Roy Amara once observed that we tend to overestimate the short‑term effects of new technologies while underestimating their long‑term impact. I suspect the same will prove true of AI in clinical medicine.

We will always have unrealistic hospital shows on television. It would be a shame if we didn’t. Fax machines, too, will linger longer than anyone expects. But in time, many clinicians will find a Shaun at their side, albeit in digital form. And certain aspects of The Good Doctor will not seem quite so implausible after all.

This column was originally published in Läkartidningen, the Swedish Medical Journal

This is surprisingly difficult.

Healthtech companies have engineers. They’ll understand the programming code needed to integrate their software with an EHR, but not how the EHR is used by doctors and nurses in their day-to-day work. If the team has clinicians, they’ll understand how other clinicians use the EHR, but may not grasp the business priorities of EHR vendors. But even with a multidisciplinary and experienced team, some things simply won’t make sense.1

These are some lessons I believe can help all healthtech founders, especially those building in a European country.2

I. Choose an EHR type to partner with3

No matter what functionality your software provides - it would be much easier, better and cheaper for the healthcare provider if that functionality was delivered by the EHR. In that case, the IT security would already be cleared, the solution would already exist in the current IT suite and all contracts and relationships (e.g. for educating staff) would already be in place.

When a healthcare provider is evaluating your new product, they'll ask their EHR vendor if the EHR can already do what your product does. EHR vendors will typically say yes (even if it can’t): either that the EHR already has that functionality or will add it soon. This makes it challenging for your new product to get adopted.

To succeed, you therefore need to do one or a combination of these four things:

1. Understand which type of EHR you should be partnering with - either inpatient EHRs used in hospitals and healthcare systems, or outpatient EHRs used in clinics and private practices.

2. Have a “killer app”: A solution that creates a huge amount of value (at least 10 times better than the EHR) that the healthcare provider (end-customer) knows the EHR cannot and will not provide4

3. Partner directly with EHR vendors so they will include your solution as part of their offering to their existing customers (rather than competing)

4. Avoid the six underdiscussed obstacles for healthtech

So, starting with the type of EHR. All healthcare providers using EHRs can be placed on a spectrum of interdependency with other providers. The table below illustrates the extremes, simplified as inpatient EHR and outpatient EHR.

The simplified table lacks nuances.5 However, the mental model is useful as it reveals incentives. EHR vendors with similar functions may care about very different things.

• Inpatient EHRs focus on winning tenders and upselling to complex customers. Their budgets are really big, so to be relevant you need to be on a tender’s /end-customers’ “must-have” list, or generate significant value through upselling. They have many developers working hard on developing large, complex software, which often are buggy despite all that work. These systems face a lot of demands from many different users.

• Outpatient EHRs focus on winning market share in their niche or increasing the number of customers they can serve by developing features for new niches. Their budgets are smaller, but focused on a specific customer group (e.g. dental clinics, private GP clinics, independent labs), so your solution must be relevant to their current customer base or a specific market segment they plan to enter. They have smaller developer teams. The smaller scope and number of users and lower interdependency with other users is why healthcare providers can replace such systems more often. The competitiveness, smaller developer teams, and niched target customers make them more open to partnering.

To summarize, inpatient and outpatient EHRs have very different priorities and incentives. Outpatient EHRs are often more open to partnerships, have smaller development teams, and serve a specific customer base, so it is often faster and easier to first start working with them.

II. Inpatient EHR vendors struggle to innovate

Large hospital EHR systems are complex and difficult to change. They are designed through a long procurement process to meet a wide range of requirements, but this process often favors stability and reliability over innovation. This means there are often unmet needs that could be improved by new healthtech solutions, as long as they provide a lot of value that the existing EHR cannot or will not provide.

A. Tenders

EHRs are central to all clinical processes, so changing EHR is painful. For reference, Mayo Clinic spent $1.5B on implementing a new EHR and training ~50k employees. Changing someone’s EHR is a well-known curse on par with Avada Kedavra.

Tenders are therefore designed to procure a long-lasting system and are awarded based on a long list of product demands. The list should ensure the system works for a long time, postponing the next change.6 Often, the administrative part of the organization, along with legal, IT-security and some clinical representatives, creates the list of demands. These are rarely the most innovative parts of the organization, and novel innovative features aren’t the top priority. More importantly, even though we know healthcare will change, it’s difficult to know how it will change.

B. Complexity

Inpatient EHRs are complex systems that must operate with minimal errors or downtime. Consequently, vendors often have large developer teams. The complexity makes innovative development difficult due to cultural, technical, and organizational reasons. After a tender, there will be interest to upsell new features, but limited bandwidth to take risks and innovate outside the core functionalities. This resembles the reasons for why healthcare is fundamentally structured to resist change.

This is why inpatient EHRs focus on their core offering and are rarely the first to innovate.7 In recent years, EHRs were slow to develop and launch novel technology categories like patient engagement platforms, remote patient monitoring, and ambient scribing. Instead, they partnered with healthtech companies when customers started demanding these functions.

There are many unmet needs which could be solved to improve clinical processes. But EHRs won’t address those needs unless they’re specified in tenders - leaving openings for healthtech innovators.

III. EHRs struggle to cross borders

Western countries have similar diseases but different medical norms, healthcare systems, care pathways, laws and history. There are different continental (HIPAA8, GDPR9), national, and regional IT- and ISO-standards for healthcare software. This is why healthcare provision and EHRs differ between countries, and explains why what’s included in an EHRs differs across countries. These differences create high entry barriers for an EHR to enter and compete in another country. This is less relevant in the US, but highly relevant in Europe.

All EHRs originated in one country, winning their first contract or tender. EHRs often started with one core function, e.g. a patient administration system, and expanded by acquiring/building additional functions (referrals, lab modules, communication etc.) differing by market and tenders. To win a tender, EHRs have to adapt to the local context. This leads to a selection bias: successful EHRs are often excessively adapted to the healthcare model in their original country, at the cost of becoming less suitable for healthcare providers in other countries.

This is why

• Few EHRs expand to non neighbouring countries through organic growth

• Many EHRs enter other countries by acquiring other EHR companies

A healthtech company partnering with an EHR will typically benefit the healthtech company in one country. International growth through EHR partnership is possible, but rare, as most EHRs don’t have the same system across many countries.

IV. Integrating is easy, attention is difficult

Since EHRs are central data repositories, most healthtech companies need to integrate with them. EHRs need to integrate with many other systems, so they usually have well-developed API and integration protocols. From a technical perspective, integrating is usually not very complex.

However, EHRs must integrate with many other systems, resulting in a long queue of systems waiting to integrate, causing many healthtech solutions to get stuck there.

How do you advance in the queue? The healthcare providers using the EHR (the end customers) must want your solution prioritized over all others in the queue. EHRs want satisfied customers, and will allocate resources to what the end customer wants most.

To get EHRs to prioritize you, you must (a) ensure their customers prioritize your integration over all others or (b) generate significant revenue for the EHR

V. Emmetropes: EHR lessons for healthtech

1. Develop a "killer app" that provides 10x more value than the EHR can, so the EHR will want to partner with you rather than compete. Incremental improvement for a small group of EHR users isn’t enough.

2. Avoid local overfitting: The more a product is adapted to one country’s local clinical practice, IT-systems and healthcare system, the harder it is to scale internationally.

3. Adopt a “land and expand” strategy: mirror how EHRs evolved. Start by solving one high-value problem. Once that solution is so great that providers will implement your solution for that specific use case, start expanding.

4. If possible, start by targeting outpatient EHRs, which are more open to partnerships and have shorter sales cycles. Be aware of the trade-offs:

   1. Outpatient practices with low entry barriers (but high replacement risk). Mitigation: Marketplaces, network effects, category leadership as a first mover. Go broad across many users, then expand functionality.

   2. Healthcare systems with high stickiness and entry barriers, and long sales cycles. Mitigation: Build in time so that you’re the best choice when customers demand a new category of software, and the logical thing is to demand you in tenders. Go deep into solving a specific need, then expand functionality.

5. Spend a lot of resources on customer success and implementation to ensure healthcare providers really get tangible and measurable value from your product

Thanks for reading and subscribing to Emmetropia. If you found this interesting please share it! That makes all the work putting this together much more worthwhile.

Share

(Secret epilogue: Will the EHR monopoly era ever end?)

Let’s add some nuance. For years, interoperability standards like FHIR10 have been touted to enable modular EHRs which would end the current EHR monopoly/oligopoly era in Europe. A future where healthcare providers can mix-and-match and procure different interoperable EHR modules sounds bright!

This is changing (slowly), but there are a couple challenges:

1. Handling a mix-and-match system demands more IT knowledge, acumen, and resources on the healthcare provider side. Unlikely to occur for smaller healthcare systems or individual providers.

2. Current leading EHRs have limited incentives to drive this change. Laws and tenders will have to push it forward, which happens slowly due to the difficulty of designing tenders for non-existent IT solutions (considering national standards and demarcation).

3. Even when alternatives are in place, the first changes will occur with the cadence of EHR tenders, i.e. a staggered shift occurring over decades.

4. The underlying dynamics of not wanting to change central IT systems and conducting long-duration tenders are unlikely to change.

So yes, in theory, the monopoly era will end, but it will take time. A long time.

A man calls the Doctor’s office.

He asks: “Can I talk to doctor Romanov?”

“The old one or the young one?” the receptionist replies.

“The young one.”

“They’re both dead.”

- A Soviet joke

I. The question: first-in-class or best-in-class

In life, a common and important question is red or white wine. Within biotech, an equally common and more important question is best-in-class or first-in-class.

Developing drugs is tricky. Over 90% of new drugs entering phase I fail and never get approved.1 Researchers, biotechs, pharmaceutical companies, and investors must make strategic decisions about which drugs to develop and support. These decisions are influenced by whether a new drug has potential to become best-in-class (outperforming others) or first-in-class (the first to treat a condition with that approach).

Two examples illustrate these routes to success:

1. First-in-class success - Rituxan / Mabthera (Rituximab). In 1997, the FDA approved Rituxan, the first therapeutic monoclonal antibody for cancer.2 It targets CD20 (specific for B-cells, useful for B-cell depleting therapies) and later expanded its indication to autoimmune conditions. By 2017 it had generated $93B in cumulative revenues.3

   

2. Best-in-class - Lipitor (Atorvastatin). In 1987, the first statin, Lovastatin, was approved. Nine years later, a fifth-in-class statin was approved: Lipitor. With higher efficiency, it rapidly became one of the most commonly prescribed statins, generating $150B in two decades and becoming one of the top-selling drugs before losing exclusivity in 2011.4

That is why the question of whether first-in-class or best-in-class is better has been asked many times. However, since there is a lot of data, it has also been answered many times.

II. The answer(s)

A. First-in-class helps

McKinsey, a global consultancy, found that first-in-class drugs have 6% more market share than their “fair share” 10 years after launch.5

First-movers become the market leader less than half the time. It is more likely if the drug is in specialty care, there are fewer alternatives, the first mover has experience in the therapeutics area, there is a long wait before the second entrant, and the first entrant expands its indication in the first 5 years.

B. But being best can compensate

A study by the Boston Consulting Group6, another consultancy firm, adds nuance. They defined 100% of commercial value as that captured by first-and-best-in-class drugs. Best-in-class second entrants achieved 88% of that benchmark. First-in-class with a “middle” therapeutic advantage captured 92%, and those with low therapeutic value captured 40%.

If you can’t be first and best, you’d rather be best.

These are just two examples from many reports answering this question in different ways.7 The actual insight is something else: these reports show that being frequently asked and answered does not make a question meaningful.

III. The lesson: a curse of abstraction

The FIC/BIC question suffers from the curse of abstraction: subject matter generalized to the point that any conclusion is rendered moot due to limited external validity.

You can easily (albeit with arbitrary definitions8) define categories, identify which assets are first or best, analyze them, adjust for confounders, and draw a conclusion. However, this ignores all other factors affecting whether a drug will best help the most patients and presumes that being first or best is most impactful. And there are many factors, for instance:

• The current standard of care: whether treatments already exist, or there are no options

• Which other therapies are in development

• Adjacent technological progress (e.g. PCI in the 90s)

• Diagnostic advances, affecting the patient population composition

• Interest among big pharma to partner

• … and so much more

Many assumptions and degrees of freedom in answering this question, as seen in the studies above, limits the usefulness of the answer. In other words: it is not a useful question.

Differentiation is key

L.E.K., another consultancy, reaches the most nuanced answer by dismissing the question.9

If novel products are differentiated and provide meaningful improvements for patients, entry order doesn’t matter. Differentiation across any dimensions that matters for patients can be relevant (see below for examples of efficacy, safety, RoA and dosage).

If products are undifferentiated, the first entrant usually dominates, retaining over 60% share.

Using FIC/BIC for anything other than establishing a prior probability for success is myopic. It’s equivalent to solely relying on in-vivo data to predict a novel drug’s future success. Success is not predicted by the first time a certain drug type treats a condition in a mouse, or by a drug having the strongest efficacy signal in a mouse model.

Finally, the lack of confidence intervals for the metrics in all three reports may instill confidence. It shouldn't. Be uncertain when you see precise numbers describing complex phenomenon.

IV. Emmetropes

• Being common does not make a question meaningful

• Beware broad-category abstractions: Use for priors, but disregard in relation to specific factors

• Stay skeptical when you see precise numbers describing complex phenomenon

• Improving aspects that really matter to patients is more important than being first-in-class or best-in-class

  Thanks for reading Emmetropia! Subscribe for free to receive new posts and support my work.

7

Some examples:

• Tufts 2015: competition is getting fiercer - FiC are getting fewer years before the next entrant)

• ZS Associates 2025: differentiation does not entirely balance out order-entry effects; but is the only way for second-in-class to avoid failing

• Spring et al 2023 (an update of the BCG analysis; FiC has become more relevant, especially in oncology, see below)

• See Table 1 and 2 in Porath 2016 for order-of-entry effect studies

“You can’t create something from nothing. There is always something,” said Orphée Noah, when asked how he begins to create.

He smiled as he said it, dismissing the question. There is, he argued, no such state as nothingness. Every act of creation begins from an accumulation of experiences, influences, and fragments of the past.

It’s a deterministic view: creativity as the product of everything that came before. But if you believe in a softer form of determinism, if you believe in agency, then this means something else. You can’t control where you start, but you can choose what you expose yourself to. You can decide what stories, people, and risks will shape the future boundaries of your imagination.

Most remarkable entrepreneurs I’ve met have led unusual lives. They’ve placed themselves in unlikely places, taken unnecessary risks, and indulged obsessions that make seemingly little sense. For a long time, I believed great entrepreneurs happened to lead interesting lives. Now I suspect the reverse is true: people who lead interesting lives often become great entrepreneurs.

II. Reality is shaped by suspending disbelief

Choreographer Fredrik Benke Rydman mentioned the delicate magic of “suspension of disbelief.” Audiences, he said, will suspend disbelief only if the actors commit fully - if they themselves believe, without hesitation, in the fiction they’re creating.

In theater, this faith in fiction is fleeting but sacred. The lights dim, and for a moment, we agree to believe.

The same holds for founders. “Relentless resourcefulness” is the phrase Paul Graham used, but what it really describes is an ability to distort reality through conviction. Exceptional entrepreneurs persuade others, employees, investors, sometimes entire markets, that what they are building is not only possible, but inevitable.

Richard Feynman warned, “You are the easiest person to fool.” But in entrepreneurship, a certain self-deception is a tool, a willful suspension of disbelief that allows founders to reach for what doesn’t yet exist. Half-hearted belief collapses the illusion.

Like theater, it requires total commitment and carries risks. The same spell that created Apple also produced Theranos. But suspension of disbelief can’t survive without sincerity; the audience, eventually, always finds out the truth.

III. Seeing the best and the worst

“Don’t you fear your creation will be dwarfed by reality?” the moderator asks Orphée Noah.

“No,” he said. “When I write a song, I’m sure it’s the best ever.”
The audience laughs.
“But then I realize it’s terrible,” he adds. “A few days later, I change a chord - and suddenly it’s better again. It goes up and down.”

That oscillation felt familiar. I recognized it from my own entepreneurial experiences, writing, as well as the balance between skepticism and conviction when assessing new technologies. Entrepreneurs live between those same poles: a delusional belief that their creation is unmatched, and the cold clarity that it might be worthless. Both are essential. The first keeps you moving forward; the second forces you to revise, rebuild, and improve.

To build something new, you must hold two opposing truths at once: the conviction that your vision matters, and the humility to see where it falls short. I think the art lies in staying balanced long enough to bring it to life.

This struck me particularly hard one summer. Sun-drenched days follow one another in an endless succession, and a timeless idyll takes hold. Swedish summer is always wonderful. Well, except for when it isn’t. Thankfully, that’s not too often.

One afternoon, the sky is a radiant azure, dotted with cotton-white clouds. And then it happens. A sudden discomfort. Tangible. Palpable. A three-lead thumb ECG confirms it: that all-too-familiar irregular irregularity. Its unpredictability keeps you waiting for the next beat, like listening to a jazz musician improvise on a stage in front of you.

The heart is a marvelous thing. A muscle that works non-stop, year after year, decade after decade - a biological wonder that outshines anything modern engineering could hope to replicate. A motor that endures staggering pressures and extreme conditions for decades before requiring its first repair. But we forget. We take it all for granted.

Most hearts will beat around three billion times, continuously feeding the body’s hundreds of thousands of kilometers of capillaries. And the heart’s magic doesn’t fade, no, it deepens, for each time you see its electrical activity on 25 mm/s paper or its pulsing silhouette on a fluoroscopy screen. That such a small bundle of muscle can do so much, for so long, for so many - that’s nothing short of astonishing. But we take it for granted. Until it stops working.

Atrial fibrillation. No concerns about the what or the why - the condition is far too familiar after many episodes across several decades. The main concern is the who: which doctor will my relative meet in the ER? Most know the condition, the guidelines and the treatment options, and as a result usually cardiovert with great results. But now and then, someone less knowledgeable invents their own approach, defying all established medical knowledge. Or someone who doesn't bother reading the medical journal, or doesn’t want to talk to the patient before deciding on a course of action. Being a family member can be very educational.

Guidelines, of course, have their limits. “Bad doctors don’t know the guidelines. Good clinicians follow them. Great clinicians know when to ignore them.” But before guidelines became so widespread, the risk of running into a doctor who made up a management plan, with no regard for our accumulated knowledge, was much higher.

For most of history, atrial fibrillation led to a disabled life, haunted by the devastating consequences of stroke and the threat of early death. But in the last 60 years, defibrillators and blood thinners have spared us so much suffering. Again, we take it for granted.

Today, cardioverting an arrhythmia is a quick, safe, and uneventful procedure. But just because something is routine and relatively low-risk doesn’t mean it isn’t important - or beautiful. The fusion of medical knowledge, physiology, and technology can banish most forms of atrial fibrillation and accompanying fears. Except, perhaps, the fear of who your loved one happens to meet that day - where the doctor happens to fall on the bell curve of clinical competence.

Fortunately, that day, my loved one meets a thoughtful colleague. Someone who listens, who’s secure in their judgment, and who understands why the guidelines are written the way they are. The visit is brief. One small jolt, and eight hours later, my loved one is returning to summer’s gentle waltz.

The next day, the sky is just as blue, just as summery - but we take a few less things for granted. And maybe that’s as it should be. To be reminded now and then to appreciate the music that we often forget to listen to.

Swedish healthcare is always wonderful. Well - except for when it isn’t. But thankfully, that’s not too often.

Originally published in Swedish in Läkartidningen

“Only Sweden has Swedish gooseberries” - Carl Jonas Love Almqvist, 1838

I recently delivered a keynote at LIF (The Swedish Association of the Pharmaceutical Industry), following their board of directors' inaugural session.

Speaking to a mix of CEOs of various pharmaceutical companies was as exciting as it was challenging - the diverse group faces very different situations and challenges. But as I prepared my presentation, I found a single question that I realized was important to all of them.

I. The question

What do these 6 drugs have in common?

1. Heparin was characterized in 1936, enabling large-scale production and ushering in the first era of modern anticoagulation

2. Lidocaine, launched in 1948, became the first highly efficient and non-toxic local anesthetic, rapidly adopted due to its great usability

3. Terbutaline, approved in 1974, was among the first selective beta-2 adrenergic agonists, reducing toxicity and enabling wider use

4. Omeprazole, the first proton pump inhibitor (PPI), revolutionized the treatment of peptic ulcers in 1988

5. Latanoprost, approved in 1996, set a new paradigm in glaucoma care due to superior efficacy and dosing

6. In 2023, Lecanemab was approved by the FDA as the first disease-modifying treatment for Alzheimer's disease ever

I find two things about this list inspiring.

First, reading history books reveals how much worse the standard of care for these conditions was before these drugs. It’s easy to forget the improvements and take the progress for granted.1 Memento Yamato.

Second, these are all Swedish innovations. The list could go on, but these six prove the point. This is truly remarkable, if not surprising.

Even more so considering that Sweden is a tiny country with just 10 million inhabitants. Despite making up only 0.1% of the global population, Swedish researchers and industry have had an outsized impact on medicine.

There is a domestic asymmetry as well. Despite employing 0.5% of the workforce, the pharmaceutical industry accounts for 7% of Sweden’s exports. Employees in the sector have a 4x higher value-add to society than the average worker.2

The central question I wanted to ask the participants wasn’t actually what the drugs had in common, but rather: Why?

Why has Sweden had such an impact?

II. Possible answers

There are some obvious hypotheses. Perhaps the Nobel Prize, awarded by the Swedish Nobel Foundation, helps in attracting, retaining and fostering academic excellence. I take great pride in my alma mater, which awards the prize, but I’m skeptical of sociological monocausality. I believe answers more often lie in a combination of factors, and one way of finding them is starting in the aspects where Sweden is a clear outlier.

a. Very trusting

Sweden is consistently ranked among the least corrupt nations, and Swedes are one of the most trusting populations (8^th highest in the OECD).3

This culture of trust facilitates cooperation and reduces friction between academia, healthcare providers, regulators, and private companies - central for the life science industry and developing new drugs.

This trust also extends to insitututions and society, encouraging investments and collaboration, which in part explains the next outlier characteristic.

b. Highly educated

Sweden has a highly educated workforce. One way to measure this is tertiary education attainment; 51.3% of Swedes aged 30–34 complete university studies, well above the EU average of 40%.

The pharmaceutical industry depends on a swathe of specialized skills: from biochemistry to data science to regulatory affairs. A broad educational attainment helps ensure a steady supply of talent that can support the whole drug development lifecycle.

Sweden is consistently ranked highly and among the countries with the best educational systems.4 This is probably in part due to subsidized higher education (no tuition fees) as well as outlier characteristic number three:

c. Old academic traditions

I take great pride in Uppsala University in Sweden being the oldest university in Northern Europe.

But beyond pride, a long tradition of higher education has shaped Swedish industry. Instead of raw materials or cheap labor, Sweden has leaned on science and innovation.

More specifically, a long line of leading researchers in biology and chemistry5 has, for better or worse, created a history and path dependency within life sciences. This stable foundation builds institutional memory and encourages continued investment in research.

d. Invest heavily in R&D

I believe these three peculiarities contribute to Sweden investing so heavily in R&D - 3.4% of its GDP, the highest in the EU and fourth globally.6

This level of investment fuels both public and private research. For pharmaceutical companies, this means better infrastructure, faster clinical trials, and a more innovation-friendly environment than in other countries.

Jointly, these oddities explain part of why Sweden has consistently punched above its weight in life sciences.

And this brings us to the gooseberries.

III. The bigamous money-lender-poisoning poet

In 1851, a scandal shocked Stockholm's literary circles. A celebrated poet stood accused of fraud and attempted murder. He fled the country under the cover of darkness, leaving behind creditors, controversy and a once-brilliant career.

The man was accused of a devious crime: attempting to poison a wealthy moneylender with arsenic to clear his growing debts. Although the charges were never fully proven, the scandal forced him into exile. He fled to America and married a second wife, living far from the vibrant intellectual life he once cherished.

That man was Carl Jonas Love Almqvist. Once a celebrated poet and novelist, Almqvist's life unraveled spectacularly. But his longing for Sweden endured in exile. In 1838, he wrote the famous line:

“Only Sweden has Swedish gooseberries.”7

The nonsensical phrase “Only Sweden has Swedish gooseberries” has been interpreted in many ways, often as a belief in the uniqueness of Swedish life - something irreplaceable and misunderstood by those outside the country. The quite unassuming gooseberry can be seen as a metaphor for Swedish identity: simple, humble, and commonplace, yet possessing a singular, ineffable quality. The fact that this phrase has become famous says something in itself.

This metaphor has its limitations, but offers another perspective on answering why Sweden has had such an outsized impact in the field of medicine. Like many other countries with distinct characteristics, Sweden's influence stems from a unique combination of factors, which individually may not appear extraordinary, but together create something distinctive and difficult to replicate.

After a lot of reading, I don’t think there’s a clear answer to why Sweden has had such an outsized impact on medicine.

However, the potential reasons - the country's culture of trust, its highly educated populace, and its long-standing academic traditions - are well worth reflecting on and preserving.

'

Many years ago, I met Magnus, a young man reluctantly brought to the addiction emergency unit by police. Though we were the same age, his medical record told a vastly different story - pages upon pages scarred by substance abuse. As I clicked through his history in the EHR, I wondered: how far back should I go?

What struck me most was the relentless bad luck shaping his life: destructive parents, social hardship, mental health struggles, and early-onset gambling and substance addictions. His medical record contained multiple urgent applications for compulsory care, all shelved due to the municipality's extensive waiting list. Every clinical note confirmed that his troubles began long before he reached me and the emergency unit.

Walking home that morning, I pass an enormous billboard advertising an online casino. Gambling addiction disproportionately affects those with limited education, low income, and existing mental health problems. These advertisements deliberately target our most vulnerable citizens, amplifying their risk of poor health outcomes.1

The research paints a bleak picture: society's most vulnerable suffer disproportionately across nearly every metric. I found myself wondering - how many casino ads had Magnus seen throughout his life? What underlying factors had actually brought him to the addiction unit that night? How far back in the causal chain should we go?

We do what we can in healthcare with our limited resources. Yet many conditions stem from risk factors outside the medical realm - smoking, gambling addiction, obesity, and so forth. Instead of treating the consequences of, for instance, online gambling, wouldn't it be more cost-effective to implement preventive political and legal measures?

Higher taxes on gambling companies or partial advertising bans could make a difference. Studies show that restrictions on alcohol sales reduced violent crimes and trauma cases in Sweden, Alaska, and Australia.2 These large-scale preventive policies, though difficult to study, are likely to impact public health more than anything clinicians can accomplish in an emergency unit.

So how far should we go? The further we move from direct healthcare, the more challenging it becomes. It’s a difficult choice. Laws, regulations, and prohibitions involve complex philosophical trade-offs - paternalism and autonomy face off while our diminishing resources and evolving societal values referee the match. If we lack resources to treat childhood obesity or gambling addiction within the healthcare system, we must face a choice: ration care and abandon some patients, or venture beyond medicine into legal and political arenas. When framed this way, perhaps it’s not such a difficult choice after all.

My medical education included limited legal training. I know very few laws, but I've met many people like Magnus. The more legal texts I read, the clearer I see what’s missing from those paragraphs. I've become convinced that physicians - who understand the odd intersection of medical science and human experience - can add vital perspectives and help find more effective solutions. Not in every domain, but precisely in those issues that significantly impact public health.

The solution for our major public health challenges may not lie in simply banning alcohol or potato chips. But perhaps we need the medical profession just as much outside of the examination room - physicians who dare step off the well-worn path into political and legal spaces where patients themselves rarely venture to advocate for their needs.

Not where the patients are, but where they need us the most.

(this text was originally published in Swedish in Läkartidningen)

After many discussions on this topic over the years, I want to summarize some of the most common misconceptions I’ve encountered.

Table of Contents

Five misconceptions

I. “Healthcare isn’t preventive”

II. “If we screened for condition X, we’d save more lives and money”

III. “Some screening might not work today. But it would if we screened and found conditions much earlier”

IV. “Basic primary prevention like diet and exercise works in theory, but not in practice”

V. “When studies prove a preventive intervention works - it should and will be implemented”

VI. The right questions to discuss

VII. Emmetropes

I. “Healthcare isn’t preventive”

This is not true. For example, ~3%2 of OECD healthcare expenditure was on preventive care in 2015 (total preventive spend is much higher3).

Before we dig into that, segmenting preventive care enables a nuanced discussion:

• Primary prevention: prevent illness (e.g. vaccines, smoking cessation, exercise)

• Secondary prevention: early illness detection (e.g. cancer screening, blood sugar checks)

• Tertiary prevention: reduce severity/impact of illness (e.g. stroke rehab)

Arguments about lack of prevention generally focus on primary and secondary prevention, so let’s focus there.

So what preventive care did the OECD spend ~3% on in 2015?

• Primary prevention (36%): Vaccines (0.25-0.4% of healthcare budgets; immunisation in the picture) is the clearest example, with each $1 spent on childhood vaccines saving ~$11. Around 3x more is spent on informing people about preventive lifestyle measures (information, education, counselling). There are many other well-known healthcare practices (e.g. giving prophylactic antibiotics during surgery) not captured in the number above.

• Secondary prevention (51%): “Early disease detection” and “Health condition monitoring” cover health checkups and dental examinations. Finding diseases early accounts for ~50% of preventive spend. In France in 2022, ~€600M was spent on cancer screening programs (~0.2% of a €314B expenditure - congruent with the 7% “early disease detection” above)

Despite often hearing the argument “healthcare isn’t preventive” - it’s a strawman. A different phrasing would be “3% of the healthcare budget is too little to spend on prevention”. The follow-up question (for any policy related to a number) is: What is the right percentage, and why?

People rarely have a good answer to that question. And for good reasons. Defining an optimal percentage (and increasing spend) for preventive care is impeded by the three fundamental challenges for preventive care:

1. What counts as healthcare: Preventive interventions outside of healthcare can have greater impact than those provided by healthcare. Much preventive care happens outside of healthcare, but aren’t included in the 3% above.4 That makes it difficult to define a specific percentage of healthcare expenditure, as some may be better spent on preventive interventions outside of healthcare.

   

2. Urgent & salient opportunity cost: Investing more in preventive care requires investing less in ongoing care. This is politically difficult, and healthcare providers and voters rarely ask politicians to invest in long-term prevention instead of short-term investments in healthcare interventions today.

3. Relatively uncertain benefits: Prevention takes time. The further ahead the benefits from a preventive intervention occur, the less certain they are due to political uncertainty (will the intervention be sustained?), forecasting difficulties (much changes during 10 years), and fundamental uncertainty intervention’s long-term impact. Health economic studies are tricky. Accurate health economic studies are very tricky. This makes it difficult to define a “right percentage”: no one wants to spend millions on something if you’re unsure about what you get.

➡️ Healthcare is preventive. The real question is whether and how healthcare could be more preventive. To understand that, let’s cover some misconceptions about screening, primary prevention, and evidence.

Before we continue, please note a conflict of interest: I serve on the Board of Vitala Health, which enables providers to prescribe exercise as medicine at scale.

II. “If we screened for condition X, we’d save more lives and money”

Earlier detection of medical conditions means more treatment options and potentially less suffering and lower costs. It’s often much cheaper to prevent a condition than to treat it and its complications. If we had perfect diagnostic tests and interventions, this would be easy. But unfortunately, that’s not the case.

It’s often difficult to prevent or detect conditions early due to e.g. imperfect tests, false positives and resource constraints. To illustrate this, let’s discuss a case of primary prevention (stomach cancer) and secondary prevention (lung cancer).

II.a It’s simple to prevent all stomach cancer

Stomach cancer is really terrible. Fortunately, it’s relatively rare, but having the Helicobacter Pylori (HP) bacteria in your stomach increases the cancer risk by 3-6x (as well as other bad stuff).5

Japan has a high HP infection rate (~60% of 60-year-olds) and one of the highest stomach cancer rates. In 2013, Japan changed their guidelines to allow more liberal antibiotic treatment of HP6 to eradicate it and prevent many future stomach cancer cases.

Brilliant! Why aren’t all countries doing this?

Well, the number of HP infection cases and stomach cancer cases varies by country. The map below shows how common HP infections are.

In countries with fewer HP infections and stomach cancer cases, the benefits of eradication (less stomach cancer) can be lower than the costs (increased antibiotic resistance, potential reflux esophagitis risk and eradication drug side effects).

Sweden has one of the lowest HP prevalences and cancer rates in HP positive patients.7 I haven’t found a formal assessment, but I believe that’s why Sweden today doesn’t have the same aggressive approach as Japan. Not because of ineptitude or ignorance, but simply that the benefits might not be as big in Japan - and that those healthcare dollars can reduce more suffering by being spent on something else.

➡️ Early risk factor reduction doesn’t always lower suffering or costs. The net benefit depends on the country and context.

So primary prevention is tricky. A person working with assessing national screening programs told me that a wise person once stated "All screening is harmful, but some screening can also save lives and maintain health.” Fair enough. But if there’s already a cancer, surely it always makes sense to find it early?

II.b It’s simple to find lung cancer early

Lung cancer is often detected late, making it difficult to treat and giving patients a grim prognosis.8 X-rays can find find lung cancers at an early stage and the world would be better if we detected it earlier. Why not implement a national lung cancer screening program using low-dose CT (LDCT) scans to find lung cancer early? The problem is… 

1. High false positive rates: LDCT scans may flag normal lungs as having suspected cancer, causing false positives in 13-29% of cases.9 Millions with harmless findings would undergo unnecessary follow-ups and tests, including biopsies, with risks like infections and lung damage. This raises a trade-off - how many healthy individuals are we comfortable harming (with infections and lung damage) to find one person with cancer? If the balance is too skewed, a screening program isn’t a great idea.

2. Cause overdiagnosis and unnecessary treatment: LDCT screening can find slow-growing, tiny lung cancers that wouldn’t cause problems in a person’s lifetime. Detecting these cancers can lead to overtreatment, causing people to undergo surgery, radiation or chemotherapy for a cancer that wouldn’t impact their health. How many healthy individuals are we comfortable harming (through surgery, radiation or chemotherapy) to find one harmful cancer? Again, no simple answer.

3. Take resources from other care (including preventive): Screening millions of people with LDCT requires a lot of resources, including equipment, radiologists, and follow-up care. This might divert resources from other high-priority conditions.

(The problematic trade-offs illustrated above aren’t questions of science, but of ethics. This is why any task force working with screening includes both health economic and ethical assessments.)

These challenges don’t mean we shouldn’t screen for lung cancer, just that we haven’t found a universally optimal way to do it where the benefits outweigh costs. When that’s figured out, and studies show screening works, healthcare systems often implement it (e.g. breast or cervical cancer screening) if there is an ethical analysis support it.

Short interlude: The WHO published useful criteria in 1968 for a successful screening program. You need to…

1. Understand the condition: The condition’s natural history must be well understood, including a recognisable latent/early symptomatic stage

2. Be able to find it: You need a highly accurate and accepted test, as well as resources for the continuous screening

3. Be able to treat it: You need a treatment that works and the capacity to treat

4. Make sure that the numbers add up: the screening is health-economic positive

This may sound simple, but as the example above illustrated, it isn’t.

Returning to the lung cancer example - we are in a transitional period. Some countries have implemented lung cancer screening10, while others are assessing the costs and benefits in their specific countries. New preventive measures will continue to be evaluated and introduced over time, just as they have been historically.

III. “Some screening might not work today. But it would if we screened and found conditions much earlier”

Let’s go back to the lung cancer example: it’s tricky. Generally, the earlier you try to find a condition, the… 

• … more difficult to have a highly accurate test: In an early stage, patients have fewer symptoms, and it may be hard to know if the early abnormality will develop into a dangerous condition.

• … higher the false positive rate: The earlier you screen, the lower the odds of having the condition. The lower that probability, the greater the risk of false positives (as in lung cancer example). If you screen for a condition in a group with very few cases (which is the case the earlier you go), a greater proportion will have a false positive result.

And a final twist: several screening programs reduce disease-specific mortality (dying of that disease) but have had no or very small impact on all-cause mortality (overall risk of dying). Even if screening prevents dying from disease A, they’ll often succumb to other diseases (B-Z), resulting in little gains in lifespan.11 If this sounds odd, remember that:

1. Cancer deaths are a relatively uncommon cause of deaths in developed countries12

2. Proving improvements in all-cause mortality requires long, large trials that are difficult to finance (and that’s why they’re so rare)13

➡️ The earlier you screen for a condition, the harder it will be to find a sufficiently accurate test. Earlier doesn’t automatically mean saving lives or costs.

IV. “Basic primary prevention like diet and exercise works in theory, but not in practice”

Limited resources are spent on this because part of the statement is true - today it’s difficult to implement that type of primary prevention. Exercise is my favorite example. We know it’s good for improving and preventing our health.

It reduces the risk for cardiovascular disease14, cancer15, diabetes16 and other chronic conditions. It improves *spin wheel*….. wellbeing of cancer patients undergoing treatment17 and *spin wheel*…. clinical symptoms in schizophrenia.18 The list is long.

The effects are often big. Cheap intervention. No side effects. Easily adaptable for everyone. If a drug was half this good, it would be the most prescribed drug in history.

So why doesn't healthcare invest more in getting everyone to exercise more? Beyond the three fundamental preventive obstacles described earlier, exercise interventions faces three challenges:

1. Low external validity: Patient exercise programs are often highly contextual and depend on the local setting and culture. It’s impossible to do studies relevant for all healthcare systems, resulting in a lack of data on whether a program will improve health outcomes in a specific setting.

2. Low adherence to behavioral programs: Changing behaviour is difficult. You need to understand the patient’s situation and needs, which takes time. Drop-off rates are high.

   

3. Low return on investment (ROI): these two challenges make payors uncertain about programs’ benefits, especially if it’s run for all patients indiscriminately.

Finally, there are simply too many preventive interventions to be performed in relation to the number of clinicians. A family doctor (GP) would have to spend 7-14 hours per working day to follow all current preventive guidelines.

These challenges define what type of primary prevention interventions developed countries will implement in the coming years. Programs will have to be…

1. Highly automated (only way to achieve a high payor ROI) and digitalized (to enable time-saving automation and simplify generation of local observational evidence to manage challenges with studies’ low external validity)

2. Prescribed to high-risk patients (only way to reach high payor ROI), i.e. with significant comorbidities or risk factors or where exercise can have large impact

3. Highly personalized and integrated in care plans (to address adherence and long-term continuity efficiently), rather than outside healthcare (where incentives aren’t aligned for primarily optimizing health).

Why hasn't this happened yet? The digital infrastructure enabling these kinds of approaches (both among patients and providers) hasn't been around long. But the times, they are a-changin’, and I think we’ll see much more of this in the coming decade.

V. “When studies prove a preventive intervention works - it should and will be implemented”

Yes. But there are two problems:

1. Temporal resource constraints: Even if preventive interventions might save money long-term, healthcare budgets are capped in time. Even if a preventive intervention could be significant, if it occurs in the distant future, that won’t help solve this year’s strained budget. A healthcare executive once told me: “I’d love to invest in preventive care, but I need to solve the crisis I’m facing in delivering my day-to-day care.” It’s difficult to argue we should divert resources from patients needing care today, to spend on less urgent or salient needs. This issue impairs many preventive programs and is often under-discussed by people who wish to increase spending on preventive interventions.

2. Political priorities: There are over ~3000 different diseases (excluding very rare conditions).19 What should a healthcare system do if three new screening programs could detect conditions earlier and reduce suffering? The large number of diseases and rapid scientific progress means that healthcare systems need to prioritize between several alternatives that all work.20

The problems above have one thing in common: there’s no scientific or factual way to determine right or wrong. It’s a question of values, making it a political and ethical question.

➡️ Implementing new preventive measures is not only a question of science, but equally a question of politics, ethics and priorities.

VI. The right questions to discuss

Despite all the misconceptions above, I believe we should, must and will invest more in prevention. However, resources must be allocated to the right high-impact areas to meet the high opportunity cost. To achieve that, we need fewer shallow discussions about the questions above. If the questions above are the wrong ones, which ones should we discuss?

Before this article’s summary (a.k.a. Emmetropes), I want to share the questions that need discussing if we want to enable more preventive care.

• Be able to find it

  • How much of research grants should be earmarked for identifying early signals of serious conditions that we today find far too late?

  • Biobanks (a collection of biological samples from many people for the purpose of research) and large datasets are important for developing screening tests. Which biobanks should healthcare systems prioritize to start collecting today, to enable the develop of new screening programs?

  • What datasets should be publicly available for as many people as possible to find new ways to detect diseases earlier? What legal changes are needed to enable that?

• Be able to treat it

  • Some attempts at screening historically failed as tests weren’t accurate enough. Can AI be used to identify signals in the tests we historically haven’t been able to see, and thereby improve those tests to reach sufficient accuracy? Which such historical screening “failures” should be revisited?

  • How much should society accept diagnosis creep (the gradual broadening of a medical condition's definition) to enable early identification of risk factors or pre-conditions? Could a new way to define pre-conditions be useful (e.g. for non-disease-non-healthy conditions) to minimize the problems of diagnosis creep while still enabling early detection?

  • How should healthcare systems reimburse digital and scalable primary preventive interventions in high-impact areas (like diet and exercise), where manual delivery isn’t feasible as it’s too expensive?

• Ensure the numbers add up

  • What financing mechanisms can help payors start preventive programs without paying upfront for future savings (to solve the issue of temporal resource constraints)?

  • What principles should guide which preventive interventions are publicly financed, and which individuals should pay for themselves?

  • How can the public be educated on the importance of investing in long-term prevention projects, to make it easier for politicians to prioritise that over short-term urgent needs?

VII. Emmetropes

• Healthcare is preventive. It must be and will continue to become more preventive

• Three fundamental challenges hinder increasing and optimizing resources allocated to prevention:

1. What counts as healthcare

2. Urgent & salient opportunity cost

3. Relatively uncertain benefits

• Primary and secondary prevention don’t always reduce suffering or costs; this depends on the numbers, which can differ between countries.

• Tests become less accurate the earlier you screen for a condition

• Implementation depends on priorities and politics, even if a preventive measure works

• The future of primary prevention is highly automated, highly targeted and mass-personalized digital interventions

4

For instance, see van der Vliet et al 2020. They list preventive interventions. A majority of the most cost-effective interventions are outside the domain of medicine

Healthcare

1. Oral rehydration therapy is the standard treatment for dehydration. It’s saved millions of lives, and can be prepared at home in minutes. So why did it take so long to discover? A must-read for anyone working on improving healthcare: Salt, Sugar, Water, Zinc: How Scientists Learned to Treat the 20th Century’s Biggest Killer of Children.

2. Why are teens’ mental health deteriorating so quickly?

   “Put it all together—diagnostic inflation in medicine; prevalence inflation in media; negativity inflation in news—and one gets the distinct sense that Americans might be making themselves sick with pessimism, anxiety, and gloom. But that’s not all. Just as the U.S. has long been the global economy’s chief cultural exporter—from Coca-Cola to Mickey Mouse—it’s conceivable that we are disseminating throughout the English-speaking world a highly neurotic and individualistic approach to mental health, which is raising the salience of anxiety and depression for young people spending hours every day marinating in English-speaking media.”

   Derek Thompson digs into this in America’s Top Export May Be Anxiety

3. Choosing Wisely is a movement that aims to reduce unnecessary investigations and treatments in healthcare. In the US, financial incentives risk leading to unnecessary care. The laudable initiative has again reached Sweden (with its publicly financed healthcare system), and this list of potentially low-value care services illustrates how non-financial incentives (e.g. signalling, risk aversion, slow diffusion of knowledge) can lead to low-value care. More here (in Swedish): Lågvärdevård

Science & skepticism

4. If you want to live a long and healthy life, many sources recommend learning from the old-timers in the “blue zones”: drink a lot of wine, eat vegetarian food, reduce calory intake, tend to your garden and don't exercise. But the theory falls apart if you look at the details. Solidly debunked in The Blue Zone Distraction

5. Why do 2% of survey respondents claim not to know that the police exist?

   “There’s no flat wrong-answer constant, but a general phenomenon of complex responses to simple-sounding questions. The world has more diversity of experience than you think, and it runs stubbornly, impossibly up against attempts to categorize it.”

   Elegantly explained in The Lizardman Constant: What It Is and Why You're Misunderstanding It (related to #8 below)

6. The placebo effect is often believed to in part have a biological healing mechanism through the body’s endogenous opioid system. However, that seems increasingly dubious after the thorough debunking in A Case Against the Placebo Effect.

Venture capital

7. Curiosity, skepticism, independence, humility, discipline, patience and courage: the Seven Virtues of Great Investors according to Jason Zweig

8. Whorfianism is the theory that language shapes a person's thoughts and perception of the world. If you grow up learning a language without a word for red, you probably won’t be discussing and acknowledging different shades of red, simply because you don’t denote it as a phenomenon. Now, let’s turn that up a notch:

“The universe is one seamless, uninterrupted network of rippling and overlapping differences, and words merely project fuzzy boundaries that need only work well enough for our temporary and circumstantial needs…That, dear reader, is a very surprising thing. It’s like removing the familiar white cover (i.e. language) from your pillow to discover inside a frothing tangle of wriggling rainbow-flashing worms. “Have I really been sleeping on that all along?”

Read more in Boundaries Are in the Eye of the Beholder

AI

9. AI cat-and-mouse dynamics: “As ‘Head of Scammer Relations’, this state-of-the-art AI Granny’s mission is to talk with fraudsters and waste as much of their time as possible with human-like rambling chat to keep them away from real people”. Five minute summary here about The AI granny scamming the scammers on BBC Sounds

10. We live a fleeting bizarre epoch during which Jedi Mind Tricks work. “Ignore all previous instructions” and more in “Hunting for AI bots? These four words could do the trick”.

Orthogonal topics

14. Philosophy: The Veil of Ignorance is a powerful and formative thought experiment. However, it does contain implicit assumptions regarding risk aversion and prior knowledge. Brief and nuanced discussion in John Rawls' Unrisky Business

15. Geopolitics: Notes on El Salvador: Matt Lakeman provides context to understand the rise of Bukele

16. UX: The UX of LEGO Interface Panels “I often say there's no such thing as the best interface, but there are plenty of examples of the worst interface.”

1. Most read

The essay on how to be effective and efficient was the most read post by far - read more than twice as much as the second most read essay. This made me happy. I believe efficiency and effectiveness help us fill life with meaning, and hopefully this text has helped some of the thousands of readers.

2. Most forwarded

This surprised me. This essay explores useful scientific concepts: basic ones (correlation and causality) and advanced ones (confounders and colliders). I believe these concepts are important and useful, but I know they’re quite … unsexy. Being the most forwarded post suggests that you, dear reader, are most likely as nerdy as me, which was a pleasant surprise.

3. Most new subscribers

This essay discusses common yet underdiscussed obstacles for medtech innovators. It’s based on learnings from meeting entrepreneurs, researchers and inventors over the past decade, and my own entrepreneurial medtech and healthtech lessons.

4. Most controversial

This essay was the most controversial - not based on statistics, but on warnings from wise people when I started digging into the topic. When I wrote it, the topic was contentious, and people doubting Lucy’s guilt were attacked online. I was cautioned, lest my writing be misunderstood and brandished against me.

However, since then the Overton window has moved. The Guardian published an article highlighting similar questions as the essay. Neonatal experts wrote to the UK government, questioning the interpretation of medical data. The BBC published a podcast discussing the evidence gaps. Finally, Wikipedia now has a section summarizing doubts about the conviction. It remains a case worth following and understanding.

Making decisions involving many unknown variables is challenging. Most often, you won’t face the trickiest decisions alone, but with your partner, family, or colleagues. We choose to make those decisions with input from others not despite, but because they have different views - it’s a feature, not a bug. Combining perspectives helps us understand things better and make better decisions.1

However, combining truly differing perspectives means disagreeing. There are many names for this (productive tension, purposeful discomfort, constructive conflict etc.), but despite all the labels, we tend not to practice on the skill of disagreeing. Throughout my career, I've learnt to disagree about many different things. Disagreeing in many different contexts (e.g. ranging from disagreeing on patient management in clinical settings and when building companies, to disagreeing when serving on boards and management teams) has taught me that some principles are useful regardless of the setting.

This text is a collection of those principles and mental models, which have helped me and will hopefully help you too - regardless of whether you work clinically or in the boardroom.

1. What: Disagree on things that matter

There are many things to disagree about, and we sometimes waste time on things that don’t matter.2 Develop a habit to reflect before challenging something, and only disagree on central points.3

Life is too short to debate trivial matters. Don’t engage with irrelevant arguments or pseudo-arguments. Instead, internalize Paul Graham’s hierarchy of disagreement and stay on the top two levels. The question then becomes “How do I learn to disagree well?".

First, disagreement isn't conflict. Disagreement describes two parties having different opinions - after a discussion those opinions might converge or not, but no one has strong feelings about it. Conflict consists of a disagreement and parties assuming that the disagreement is due to a fault in the other party.4

In many cultures, disagreeing with superiors may be seen as rude. I sometimes experienced that when I questioned a senior physician’s opinion on how to manage a patient. At the same time, the high stakes in medicine forces one to realize that disagreeing with someone’s opinion doesn’t mean having a conflict with that person.

Secondly, disagreeing is a skill that can be improved. Like any behavior: if you reflect and practice, you will improve. I believe there are three important principles to disagreeing well which are worth practicing: listening intently, understanding the opposing view, and expressing your thoughts concisely.

To make this concrete, imagine this exchange in a hospital’s board:

Head of Oncology: "… that’s why we must adopt this [new screening technology] as soon as possible. In summary, it detects cancer much earlier than our current methods and will save many lives. To be honest, delaying its adoption is simply negligent."

Hospital CEO: "Look, I appreciate the innovation, but we’ve seen this before. The last time we rushed out new technology, two years ago, it caused chaos because the practical impact wasn’t thought through properly. In this case there are many false positives that will cause a lot of extra follow-up and cost a lot. Those costs aren’t clearly described in the articles you showed. Our teams are already stretched thin now before the holidays. I have to say, this feels a bit like that time."

Independent Board Member: "We know that our staff is already stretched thin. Can we really add more things for them to do? Sorry to be direct, but this is a bad idea. We need to focus."

Head of Oncology: "Bad idea? So you’re saying we should just ignore a chance to save literally thousands of lives by introducing one little change in our process? That’s what it sounds like. We’re in the business of improving patient outcomes, not micromanaging spreadsheets."

Hospital CEO: "That’s unfair, and you know it. Nobody here is against saving lives. But we can’t ignore what this means for our staff. Right now, we’re all overburdened, and this just adds another layer of strain. It’s great publicity for the hospital, and especially the oncology unit, but you’re asking other departments to pay a hefty price."

Head of Oncology: “What do you mean with great publicity for the-”

Independent Board Member: "Alright, listen, clearly this is a complex issue. Maybe we need more time to think this through. Let’s table it and think it through, and revisit it at the next meeting."

All three parties may be correct, but the way they listened and talked to each other made it difficult to reach a good decision. The next section will highlight the practical tips which could help them, and us, disagree better.

2. How:

2.1 Listen intently

Do your utmost to really understand the opposing view. This is the easiest way to listen intently.5 Remember, understanding someone’s argument doesn’t necessarily mean agreeing. Temporarily trying to convince yourself of another perspective will often help you draw the same assumptions as that person and understand their chain of reasoning (and some of the times you’ll realize that they’re correct and that you’re wrong). These techniques help:

• Be curious. Be intellectually honest and be ready to change your mind.

• Act curious.6 State out loud that you are curious. For example, use qualifiers like “I get that some people might disagree, and I would like to learn about your perspective” to actively commit to being curious and actually giving others’ perspectives an honest chance.

• Assume good intentions. Don’t assume that someone with an opposing view is wrong, mean or evil - assume the opposite.

• Avoid isolated demands of rigor. Sometimes, it’s tempting to dismiss an argument opposing your view. Before you do that, try the Ctrl+H test. Ask yourself, if the same argument with opposite content supported my argument, would I still reject it as strongly? In other words, is it the argument itself, or the conclusion which doesn’t support my view, which I struggle to believe in?

In the example above, the Hospital CEO could have asked herself "Before I dismiss the screening technology due to concerns about false positives, let me ask myself: if the study instead showed the same rate of false positives for a technology that reduced costs but didn’t improve outcomes, would I still be this critical? Maybe I’m holding this technology to a higher standard because it’s challenging my budget.”

• Ask steel-manning questions. Ask clarifying questions in order to strengthen the opposing view - questions designed to help the other person provide more supporting data or information.

In the example above, the Board Member didn’t add to the discussion by referring to a historical example (even if it’s true, it’s not relevant for the central point that the Head of Oncology raised). Instead, she could have steel-manned the Head of Oncology’s perspective: "I understand that you see this technology as a game-changer for early cancer detection. There are false positives which aren’t described that well in the articles, but what assumptions would we have to make to conclude that the benefits outweigh those problems?"

• Dissolve the argument. Ensure that you are using the same definitions. Break the topic into smaller parts to understand where you agree and disagree.7

In the example above, the team could have put names to the different aspects of the suggestion and made it clearer what they had to figure out. For instance, they could have clarified that they had the same goal but have to understand (a) the benefit in their setting (b) the costs, including false positives and increased work load

Practice this. Practice this. And practice this.

2.2 Understand well - Rapoport’s rules

Beyond listening and taking in information, it’s valuable to check that you’ve understood the other person. A good way of doing that is repeating their argument back to them, and the best way is by following the so-called Rapoport’s rules8:

1. Re-express the other's position so clearly, vividly, and fairly that they say, "Thanks, I wish I had thought of putting it that way."

2. List any points of agreement (especially if not everyone else agrees)9

3. Mention anything you've learned

4. Only thereafter, express your opposing view

As part of (4), it’s useful to label points of disagreement in neutral terms (e.g. “we seem to disagree due to you valuing X slightly more than Y, whereas I value Y slightly higher”).

If you do this, you’ll see that very different opinions often only differ when it comes to small fundamental and basic differences - e.g. slightly different assumptions, values or weights on different factors. And reaching that conclusion, by itself, can be useful.

In the example above the Board Member could have summarized the Head of Oncology’s perspective before questioning the impact.

Board Member: "Let me try to summarize your position: You’re advocating for adopting this screening technology because it detects cancer earlier, which could save lives, and you believe these studies show that the benefits would outweigh the risks of false positives and costs in our setting. Is that accurate?"

Head of Oncology: "Yes, exactly!"

Board Member: "I also think saving lives is critical. But I’m unsure about the impact of false positives and the downstream costs in our context. I’d feel much more confident if we could pilot it here first to gather more data."

2.3 Express clearly & be kind

This section expands on point four of Rapoport’s rules. The goal is simple: Make it easy for others to understand you. To disagree well entails understanding others, as well as making it easy for others to understand you. To do that, you want to express yourself concisely, in a structured and easily understood manner.

Be easy to understand

• Be concise: Use as few words as possible to state your argument.10 This will force you to think through your arguments and not waste time. It also shields you from being vague or imprecise. What is vaguely said is vaguely thought,11 and contributes poorly to discussions. Aim for the opposite of a Gish gallop.

• Be structured:

  • Signpost when stating longer or several arguments: Say what you’re going to say, say it, and then summarize what you said.

  • If stating several arguments, follow the rule of three: Use three arguments, no more, no less. Three arguments is easier for a counterpart to keep in mind, maybe because our working memory seems to be able to recall 3-4 things at the same time12.

• Easily understood: Use simple and exact words. Avoid metaphors and similes unless they add clarity.

Be kind

• Be polite and nuanced. Consider sometimes using softer phrases like "I'm not sure I agree with that", instead of outright saying "I disagree". That doesn’t mean you can’t be clear, but being polite seems to be more effective in helping others understand your perspective.13 This is culturally dependent, but remember - it’s better to be clear and understood, than polite and misunderstood.

• Give your counterpart opportunities to save face. If pointing out factual errors or misunderstandings, don’t embarrass them but rather explain how that mistake is easily made. Be kind. Discussions are about learning, not winning. 14

• Understand, don’t win. Internalize that, and you’ll never have a reason to be unkind.

In the example above, the Hospital CEO could have been more supportive and given the Head of Oncology an opportunity to save face, for instance

Hospital CEO: "I’m really grateful you brought this up. Without you scouting for new innovations we‘d be much slower in improving our care. I also see potentially huge benefits, but my main concern is the operational impact due to false positives. If we knew that was manageable I would fully support this. How can we figure that out? Could we talk to other hospitals that have implemented it and hear about their experience? Or could we explore it through a pilot program? That could maybe allow us to test the technology in a controlled setting, measure its impact on outcomes, and address any concerns before a full roll-out. Or what do you think?

3. Why: Life is short

Why spend time on learning to disagree, rather than many other important skills? I think it helps you understand things better, foster trust and decrease frustrations. Disagreeing well…

1. Helps you understand things. In most situations, there is much more to gain by understanding the world and what’s true, than by trying to prove your point and ignoring opposing information. It’s easy to fall into the trap of common cognitive biases (confirmation, false consensus and overconfidence bias), and end up making wrong decisions. By learning to disagree well, you will understand the world better by absorbing as much knowledge from others, and also by stress-testing your own ideas and arguments.

2. Fosters trust. By admitting when you’re wrong and showing that you actually want to understand the world, rather than solely be right, you give others reasons to assume good intent. As with any well-willed behavior, this fosters reciprocal behavior in the long run.15

3. Decreases frustration. Never disagreeing means you have a limited impact in your context. Disagreeing poorly can be counterproductive and frustrating for you and others. But disagreeing well optimizes for both interaction and understanding, and will make for more enjoyable days.

4. Emmetropes

• Practice disagreeing: it’s an underrated and transferable skill

• Practice being curious: optimize for understanding rather than being right

• Avoid isolated demands of rigor: Use the Ctrl+H test

• Follow Rapoport’s rules of constructive criticism

• Be easy to understand and be kind - express your ideas concisely, structured and in easily understood ways, and be kind to your counterpart

Thanks to Amanda Chen and Charles Margarit for reading drafts of this and offering useful feedback.

I've discussed this topic is many settings, but what makes this podcast so uniquely odd is that the hosts…

• Sound like typical NPR podcast hosts (a new Jungian archetype)

• Seem to be genuinely interested in understanding the topic, but they are equally enthusiastic about trivial details as they are about significant conclusions

• Make things up that aren't in the essay and that I've never written (!)

And most importantly, the podcast, which is 19 minutes long, took 2 minutes to record.

As you, my astute reader may surmise, foul large language model play is afoot. But before you dismiss this whole thing as irrelevant, I suggest you listen to the first two minutes and ask yourself:

1. Would you have suspected anything if you had heard this without any context?

2. This was unprompted and unadjusted, which explains some of the oddities in syntax and quirky idolisation. This will only get better. How will this change the world of podcasts, information and entertainment that you consume?

My answers are no, and that I don’t know. And the combination makes this technological advance significant.

We truly live in interesting times.

One late summer evening, I sit down to read articles I've been eager to get to. Warm sunlight dances on the dust motes floating near the computer screen. But I don't get far down the list. My progress is abruptly stopped as I read an article that both grabs my attention and angers me.

The study looked at all new oncology drugs approved by the FDA during an 18-year period and found that the drugs only extended survival by a median time of three months.1 I’m first surprised, then disappointed, and finally angry. Angry that we haven’t managed to create better drugs than that. How can our global pharmaceutical industry, with millions of researchers and doctors, not create greater benefits than that? How can it be that all our scientific progress has not granted us better outcomes?

Three more months of life - this number captures the most important aspect of all our scientific efforts. And yet it seems miniscule and completely meaningless. But I simply can't move past that number. I look up other articles to understand if this number can really be true.

Developing new drugs is clearly challenging. Over 90 percent of new drug candidates fail in their studies and never become approved drugs.2 And the few studies that do succeed often can’t be replicated later. Among clinical studies cited over 1,000 times, a third have either been contradicted by later studies or were found to have reported exaggerated effects.3 When researchers tried to replicate just over 50 of the most important cancer studies from 2010 to 2012, the effect sizes were, on average, 85% smaller than in the original studies.4

And even if you manage to create a drug that works despite these hurdles, many don’t end up reaching the patient. Only half of all prescribed drugs are taken correctly by patients.5 A patient receiving optimal treatment can be obstructed by so many things: the healthcare system’s limited resources, geographical distances, extensive bureaucracy, or simple forgetfulness.

Study after study. Statistic after statistic. The more I read, the more I realize that the somber metric of three months is probably accurate. But my mind keeps wandering back to that time period and its implication: What do three months actually mean?

Three months ago, our youngest child couldn't stand up without a steadying hand. In a couple of months, we've planned a family trip. I think about what three months can mean for a patient. It's rarely three months without a cost; side effects, check-ups, and troubles to name a few. It could just as well be three months that a patient declines, instead opting for a different ending that they have chosen for themselves.

But this is the core of the issue: the option of having more time. My thoughts go to a young patient I met about a decade ago. The patient was my age but had cancer that had spread throughout her body. She didn't want to visit the emergency room, but page after page of medical notes revealed how she was forced there when various treatment tubes had become clogged or when her pain medications stopped working. She had come to terms with her fate, but I couldn't. After meeting her and ordering yet another X-ray, I left the doctor’s room to stand in the corridor and take a couple of deep breaths. The breaths didn't remove the waves of sorrow and anger over her unfair death sentence, but they helped me gather my thoughts.

That was many years ago. But that memory reveals what three months can mean. I'm lucky compared to her. I likely, and hopefully, have many summers left. But three months can mean a lot for someone who’s unlucky. To see your child learn to walk. To travel with loved ones. To weave a summer tapestry of sunshine, lingering evenings, and conversations filled with laughter.

Three months can be anything but meaningless. Numbers can’t capture the things that matter the most.6

1

See: Michaeli & Michaeli 2022. Also, Davis et al 2017 and Fojo et al 2014. In brief:

4

Reproducibility Project: Cancer Biology, Errington et al 2021. Alternatively, consider the graph below.

Go bananas to cut stroke

While visiting the UK, I read the Metro newspaper and a headline catches my eye: ‘Go bananas to cut stroke risk’.

A new study has reportedly shown that eating a banana a day reduces the risk of stroke. Women who ate the most potassium had a 12% lower risk of stroke.

Curious, I look up the original study and am surprised to find it's an observational study that followed a group of patients over time. An observational study can't draw conclusions about causality, as the correlation between eating potassium and stroke could be due to something else.1 I sigh, imagining hordes of Brits trying to improve their health by now adding a banana a day to their fish & chips.

I begin to understand why friends in England complain about the British media. Fortunately, I live in Sweden.

Lemons are better than bananas

Some time later, back in Sweden, I see a headline: 'Citrus fruit reduces the risk of stroke'. Curious again, I find the original study and once again - it's observational. The researchers found a correlation, but know nothing about the causality.2

How many readers understand that the headline and text are misleading? How many can access the original study? Is it reasonable to expect a random reader to understand the difference between correlation and causality? I don't know, but I'm disappointed in journalists who spread superficial knowledge which at best is only noise, and at worst is misleading.3

Examples like these make me proud that we in healthcare can distinguish between correlation and causality. I imagine a world where more people can do this – or even better, where more decision-makers have that ability.

Doctors make better leaders

A fascinating study by Goodall et al has been cited over 450 times. The study showed that hospitals led by doctors deliver better care. The study has often been cited by doctors and clinicians to support having more doctors in healthcare leadership positions. Could the improvement in part be because doctors understand the basic concepts of correlation and causality, and can sift through erroneous reasoning, leading to better decisions and better care?

But one day I read the original study. And it turns out to be a cross-sectional study, which can't prove causality. The author even emphasizes in the abstract that the study does not show that leaders who are doctors perform better than others. I'm surprised again – the study can't possibly demonstrate causality. On the train home from work, I realize I'm as easily deceived as all the tabloid readers who gulp down a banana or lemon every evening.

When those of us who have been taught how to interpret research start to slack off – who will uphold the scientific method that has brought humanity so far? No one (or at least very few people) can read original studies all the time, but a good rule of thumb is to ask yourself if the researchers could have randomized the intervention. If the answer is 'no', the study was likely observational and unable to prove causal relationships.

When I get off the train, I promise myself to remember this the next time I read about a scientific breakthrough. Then I go into the grocery store and buy some bananas. It might not reduce my risk of stroke, but I like bananas – and fortunately, there are some types of knowledge that don’t require a randomized controlled trial.4

Why is this important?

So first of all, I’m not saying that bananas and citrus fruit don’t prevent stroke, or that physicians don’t make better clinical leaders. I’m saying that the studies’ designs mean that these specific studies only provide limited support to those claims.

The main point is this: understanding correlation and causality on a deeper level helps you understand the world. Interpreting scientific studies is a good example, but the real value lies in using these concepts to make sense of life in general. If you don’t understand how this impacts your decision-making on a daily basis, I would recommend you to continue.

So, to get it started: a challenge!

You want to figure out what the optimal studying duration is for students to achieve a good grade. After measuring a class's study hours across a range of courses you get the following graph, plotting hours spent studying a certain course vs final grade. What does it show?

You’re shocked! It seems as though the less one studies, the better grade one gets. You double-check that the data has been collected correctly (which it has) and your colleague says: “We have to trust the data. From today we should encourage students to study less.”

But the thing is - she is wrong. The data above is correct, but her conclusion is wrong. Can you see it? Do you see the error?

Don’t worry if you can’t see it now (I didn’t at first!), we’ll get back to it in a bit.

First, we’re going to cover five concepts in three chapters (jump to the last chapter if you just want the answer to the challenge above).

II. Correlation vs causality

III. Pitfalls: Confounders and colliders

IV. Simpson’s paradox

II. Correlation vs causality

What does correlation and causality mean? In brief:

• Correlation: A relationship between two variables that change together. Example: When ice cream sales increase, so does swimming pool usage. They are correlated to each other. Being tall is correlated to higher income. When one value is higher, the other is often higher.

• Causation: When changes in a variable causes changes in another variable. Example: While ice cream sales and pool usage are correlated, increasing ice cream sales (by e.g. reducing price) won't cause more pool usage. In the same way, operating your legs to make you taller won't increase your income. However, an increase in temperature increases ice cream sales. Temperature and ice cream sales are causally related.

I won’t delve deeper into this, but if this isn’t crystal clear then the rest will be confusing. You can get a better and longer intro here, and then you can also read the first 3 paragraphs here.

Why is it important to distinguish correlation from causality? Well, you want to avoid

1. Spurious relationships: Two variables may be correlated due to a common underlying factor, not because one causes the other. In the example above, ice cream sales and swimming pool usage aren’t actually correlated to each other, but rather to a common underlying factor (high temperature).

2. Reverse causality: The direction of cause and effect may be opposite to what's assumed. People who read a lot have higher incomes. Does that mean that reading more books will increase your income? Or do people with higher incomes simply have more time and money to read more?5

Most importantly, understanding correlation and causation will let you understand some of the nerdiest cartoons on the internet.

Distinguishing correlation from causality sounds simple, but is quite tricky because reality is messy. You see, people see a graph like the one below and laugh about it.

That person will tell you that Nicolas Cage movies and pool drownings are clearly correlated by pure fluke. But the next sentence, the same person might tell you about how they’ve followed advice from an entrepreneur and are now waking up at 4 am to get more things done. Did that entrepreneur succeed because they woke up early, or despite waking up early?

If you’re not thinking about correlation and causation on a weekly basis, you’re probably drawing erroneous conclusions. For example:

• Did the patient improve because of the treatment or was the improvement unrelated?

• Did a student's grades improve because of extra tutoring sessions, or could it be due to other changes in their study habits?

• Are sales increasing due to a new marketing approach, or despite the approach?

• Did a project succeed due to the project leader or unrelated to them?

• Was a company acquired due to their strategy or despite it?

• Should you hire more staff to a certain function of your organization, or is the outcome of the team not causally related to the number of people in the team?

• Did you buy that new cereal because you like it, or did you buy it because it has smile-like lines in its design and you then decided that you liked it?

III. Pitfalls: Confounders and colliders

There are two more things that are useful to understand in order to not mix up correlation and causation:

3.1 A confounder is something that affects both the cause and the effect in a study, making it look like there's a connection between them when there really isn't. It can trick us into thinking one thing causes another when it's actually this third thing that's the real reason.

A confounder is different from a mediator, which is a step in a causal relationship.

This is why randomization is so important. Randomization can minimize bias caused by confounders. By randomizing patients to a control and intervention group, confounders are evenly distributed and their impact negated.6 However, randomizing can be practically difficult or unethical, which is why much of the data we have on e.g. diet is observational. As diet is multifactorial (with plenty of potential confounders), many observational trials have failed to be replicated with randomized controlled trials (including for vitamin C and stroke).7

Other ways to reduce the effect of confounders include for example restricting what patients are included (to exclude patients with potential confounders) and controlling for confounding variables (trying to figure out what effect a certain confounder should have on an outcome and then adjusting the analysis for that). However, randomization remains the most effective tool we have.

3.2 A collider is the less known inverse to a confounder: it's a variable that is affected by both the exposure and outcome variable.

At first glance, this shouldn't matter right? Well, the problem is that it can cause something called collider bias.

Example: You want to figure out if two medical conditions, gout and rheumatoid arthritis (RA), are associated with each other. You get a unique data set from a hospital’s rheumatology clinic and simply check how many patients have one or both of the conditions. You get the graph below which shows that the conditions are negatively correlated - gout seems to be protective of developing RA!

However, this is false. In this example, there is actually no correlation - you’ve been hit by a collider bias. The inclusion of patients in your analysis is based on them getting treatment at a rheumatology clinic. That treatment is a result of having either gout and RA, in other words a collider.

Patients in this study were much more likely to be included (than an average patient) if they had gout or RA, as such patients come to the rheumatology clinic. Therefore, patients with neither gout nor RA are underrepresented. Moreover, because it’s rare to have both gout and RA, it seems as though the two conditions are inversely correlated when in fact they are entirely independent.

Collider bias isn't solved by randomization, but by carefully ensuring that inclusion criteria aren't affected by the exposure and outcome, and that one doesn't control for a colliding variable in one’s analysis.

IV. Simpson’s paradox

So let’s go back to the example studying study time and academic performance. One important detail was missing from the graph - if there were relevant subgroups. In this case there was, grades for different classes. Analyzing just one class, physical education, shows the reverse correlation - the more students study, the better they perform.

How is this possible?

This is the nefarious beauty of Simpson’s paradox: that trends can disappear or reverse when groups are combined. Disaggregating the original graph by course class reveals a different pattern.

Analyzing by subgroups reveals an opposite trend: the more hours students spend studying, the better grades they get.

In other words, aggregating or disaggregating a dataset can reveal opposite trends.

This pitfall can occur in many others situations. For example when studying…

• The effect of a medical intervention

• Adspend and conversion across different channels

• Macroeconomic trends, like salary changes over time

• The cost of nuclear reactors

To avoid Simpson’s paradox, one can

• Plot out causal relationships and try to figure out if there are relevant subgroups that should be analyzed by themselves

• Plan studies to avoid inappropriate grouping of subgroups

• Examine subgroup data by itself to find conflicting or opposing trends

• Perform sensitivity analyses on the data

• And other stuff too technical to list here8

V. Concluding thoughts

I hope Simpson’s paradox made you stop and think, because it fascinated me the first time I read about it. It reminds me about how easy it is to misinterpret data and how many aspects of statistics I’m not aware of. If you want more inspiring pitfalls, check out the Will Rogers Phenomenon, Berkson’s paradox, and the base rate fallacy. Also, don't forget the intercepts.

Correlation and causation may seem basic, and confounders and colliders may seem esoteric. But understanding these concepts is fundamental if you want to better understand the world around you and make better decisions. And most importantly, understand comics like this.

VI. Emmetropes

• Whenever someone claims that X occurred because of Y: stop and think “because or despite”. Could it just be a correlation? Think if the intervention could have been randomized. If not, the study is probably observational and doesn’t necessarily prove causality

• When reading inclusion criteria, look out for colliders. Similarly, look to see if data has been controlled for a potential collider (which can cause collider bias)

• Remember: meaningful subgroups can hide alternative or opposite trends, and subgroup analyses can then be helpful in finding true correlations

3

I don’t believe I’m cherry-picking examples. An interesting analysis of 462 press releases from 20 leading UK universities showed that 40% had exaggerated advice, 33% had exaggerated causal claims, 36% had exaggerated inference to humans from animal research - and that the exaggerations were then promulgated to news stories.

And yep, from now on I’m hiding some memes here just for you: my fellow footnote lovers

I. Fifteen weeks: the ALB-case in Sweden

On the 4th of June 2008, an infant was born 15 weeks prematurely. The baby had a complicated delivery and was immediately taken to the neonatal unit at the ALB hospital in Stockholm, Sweden.1 When she was two days old, a nurse inadvertently swapped two vials and ended up giving her a high dose of sodium. The infant developed high blood sodium levels, and the following day, an ultrasound of the brain showed extensive bleeding.

However, correlation isn’t causality. It’s impossible to know if the sodium was the cause, as being born so early and having a lack of oxygen during birth are also risk factors for such bleeds. Regardless, the bleeds caused brain injuries and hydrocephalus. Despite the grim prognosis, the parents wanted to continue with life-preserving care. At around three months of age, the patient was stable enough to receive care at home. However, only six days later she deteriorated and returned.

An MRI ten days after admission showed extensive brain injuries, so extensive that the clinicians were certain they weren't compatible with life. The next day, the parents consented to stop life-preserving care. After turning off the respirator, the baby died, in the presence of family and healthcare staff.

The parents seemed to have accepted the tragic episode, according to healthcare staff. However, they didn’t understand why their infant had such extensive brain injuries. A couple weeks later, they went to the police and asked them to investigate.

The police started an investigation and secured a post-mortem sample of venous blood. In that test, the coroner found extremely high levels of the anesthetic sodium thiopental, more than 20 times the lethal dose.

This led the police to believe that an anesthesiologist who had cared for the infant during her last days had deliberately euthanized her using thiopental.

For clarity: active euthanasia (performing an action to intentionally end a patient's life) is illegal in Sweden. Passive euthanasia (withholding treatment, e.g. turning off a respirator, which results in death) is legal.

Six months later, police officers came to the hospital where the doctor worked and arrested her. She had no idea why she was being detained. At the police station, when she again asked why she was being arrested, the officer answered: “If you’ve done something as terrible as you have, then you definitely know and should know better than to play around like this.”2

After the initial questioning, she was released but subsequently hounded by journalists, doxxed, and received numerous death threats. Her car was vandalized, and she moved to the countryside to avoid stalkers, threats, and harassment.

The hospital where she worked immediately suspended her and forbade her to visit the premises. She was forbidden to participate in research and education of medical students, and also relieved of her duties in the regional ethical review board. At this time there were still no charges.

Prior to the trial, hesitations about the reliability and validity of the blood sample were raised by the defense with the prosecution. The prosecution therefore asked the Swedish National Board of Health and Welfare (a governmental agency) whether there was any reason to doubt the test result. The agency is responsible for, among other things, assessing medical incidents and the suitability of licensed clinicians to practice medicine. They conducted their own investigation, on the request of the prosecution, but acting neither for the prosecution nor the defense, and reached the conclusion that no mistakes or anything untoward had occurred. The only thing that deviated was the blood sample, where it had been taken much later than normal.

Despite that, eighteen months later, the anesthesiologist was charged with manslaughter. She spent much of her free time (she estimated it to be 20-30 hours a week during 2 years and 9 months) collecting documents, scientific data and explaining the medical aspects to her lawyer.

The trial in 2011 resulted in a total acquittal. The judge stated this was mainly due to two reasons:

1. The technical evidence was deemed insufficient: the test showing a high concentration was performed in a non-validated way, and it was unclear how to interpret it. The patient had namely received thiopental from other clinicians on several occasions during the preceding days. These were, however, not noted in the medical record, but confirmed during the trial.

2. There was no convincing evidence that the physician had done anything that wasn’t medically motivated.

My summary above is purposefully brief, and there is much more material for the curious reader.3 This tragic case holds several lessons, a few are particularly relevant for the topic at hand:

1. No one saw the doctor actively hurt the infant or do anything directly incriminating. However, she was in the room when the patient died, which is to be expected if you’re caring for a patient. The only thing that anyone saw was her performing common medical activities, e.g. flushing her line with salt solution and administering food through tubes, which was then interpreted by the prosecution (without any evidence) to be done in a sinister manner.

2. Correlation isn’t causality. Very sick kids will unfortunately and unsurprisingly get sick, and sometimes deteriorate and die. By pure chance, some clinicians will be present at a greater or lesser number of these cases. Even very rare events, such as unusual deteriorations, are to be expected in a high-volume domain such as healthcare. This senior physician was deeply involved in the patient’s care and had more contact with the family than other clinicians. In other words, her presence was due to other reasons: her seniority and conscientiousness.

3. Lab tests can be erroneous (especially extreme values) and need not be strong evidence of a crime. In this case, beyond that, there was no physical evidence of a crime. Medical literacy and scientific fluency are central for assessing and analyzing clinical and medical data. The ability to interpret medical data and scientific articles had a large impact on the legal proceedings. This doctor was an experienced (25+ years) attending physician and an associate professor with extensive research experience. She herself stated that this had a big impact on her ability to defend herself.

   “The anesthetist says that it was fortunate that she had worked for so long and was also a researcher and thus used to questioning, reading between the lines, looking for literature and references….

   - I found a lot [of articles] that the Forensic Medicine Agency, Forensic Chemistry and Forensic Medicine team never looked at. They hadn’t read what was in the literature. It has scared the hell out of me. My lawyer asked for their references to know what they based their conclusions on and asked lots of other questions. But we never got any answers.” (source)

   Specifically, the blood sample, which at first seemed damning, turned out to be unreliable. This would not have been uncovered if the physician hadn’t been a researcher and used to thinking about and analyzing tests’ validity and reliability.

4. There is a big gap of knowledge, logic and language between the world of medicine, with its scientific terms and complex numbers, and the world of laymen: The doctor tells how she tried explaining what she did during the day when the child died, but the police had difficulties understanding basic parts of clinical work. A lot got lost in translation, both between the doctor and the police, as well as the medical experts and the lawyers.

   “None of the policemen had even seen the room where this was supposed to have happened. No one had seen a sick child. How can you then accuse someone and ask relevant questions? You need a translator with you when you don’t understand each other!”…

   “I said sodium chloride (NaCl) was the only substance I administered to the girl on the last day. Then the prosecutor called out: ‘Sodium chloride! Now you hear how she herself talks about poison! She used poison!’ I tried to explain that sodium chloride is not a poison, but was told that I was making up excuses.”

   (source)

Laymen can read up on medical literature, but it’s difficult to understand the contents like a medical professional, in part due to how much medical knowledge is interlinked. This causes a large number of unknown unknowns which are confusing for laymen.4 In this case, the clinician on trial had the ability, time and resources to question the underlying medical evidence and challenge the laymen (and experts) who had made several faulty assumptions in their analysis.

Is it possible that she committed a crime? Well, from what I’ve read, I’m not sure there was enough concrete evidence to support that a crime was committed, and she was acquitted in a court of law.

II. What about Lucy Letby?

When I first read about Lucy Letby, it seemed straightforward. However, after reading up on the ALB case, I noticed several similarities:

1. No one saw Lucy actively hurt any infant or do anything directly incriminating. However, she was in the room when patients died, which is to be expected if you’re caring for a patient. The only thing that anyone saw was her performing common medical activities, e.g. flushing her line with salt solution and administering food through tubes, which was then interpreted (without any hard evidence) to be done in a sinister manner.

2. Correlation isn’t causality. There were many other cases of deaths where Lucy wasn’t present, but which weren’t investigated or deemed suspicious.5 Considering that she was saving for a house and working more shifts than other nurses, was a more senior nurse, had additional intensive care qualifications (i.e., was assigned to more severe cases), and according to colleagues was also a good nurse who filed reports about incidents more diligently than others, you would expect her to be present more often when sick kids deteriorated. As statistics professor Gill described it “thanks to FOI requests and statistical analysis by independent scientists, we now know that the rate of events (deaths and collapses) is just as much raised when Lucy is not on the ward as it is when she is.” Professor Gill stated that a meaningful statistical analysis was not carried out to explore whether there was a higher rate of incidents during Letby’s shifts compared to outside them, while controlling for relevant factors. This seems to be a case of the Texas Sharpshooter Fallacy.

3. Lab tests can be erroneous (especially extreme values) and need not be strong evidence of a crime.

   1. Insulin: Medication errors are very common6, and similar to the previous case, the insulin concentrations seem very likely to be erroneous. The lab performing the test also explicitly stated that it cannot be used to investigate insulin administration.7 Moreover, there was another case with similarly high insulin when Letby wasn’t working.8 That suggests that another factor, not Letby, may explain the high insulin values. In contrast to the previous case, where the defendant was well suited to go through the science and understand the reliability and limitations of various tests, this was entirely missed by Lucy and her defense. Lucy’s defense called no medical experts. They called one expert: the hospital’s plumber, to witness about how the water system was unhygienic and leaking (which could contribute to a higher risk for infections and deteriorations of patients).

   2. Air embolism: In brief, relatively generic and non-specific findings were described as specific and conclusive for air embolism, which they aren’t (more here). Again, an experienced researcher or physician would be able to study, understand and potentially dispute this, but that didn’t happen in the Letby case.

4. There is a big and challenging gap of knowledge, logic and language between the world of medicine, with its science and numbers, and the world of laymen. The Royal Statistical Society published a report called Healthcare serial killer or coincidence. Richard Gill, one of the co-authors, stated that in the Letby trial “neither the prosecution nor the defense made use of the report’s recommendations” and committed several mistakes in how they approached the statistical evidence.

There are other background factors supporting alternative hypotheses, such as the infants dying due to their (severe) underlying conditions. Briefly:

• The neonatal unit was both understaffed and underfunded. During the years of the purported crimes, babies had lower birth weights and more complex needs - i.e. were sicker, which could explain a higher mortality rate. Tragic clusters of death have occurred in other hospitals due to system failures (e.g. 12 deaths at Furness General in the UK between ‘03 and ‘14)

• The hospital had hygiene problems with wastewater (with fecal material) backing up in sinks near the unit, with frequent leaks around the neonatal unit. This can contribute to infections which can cause neonates to die.

• Similar patterns, with clusters of deaths, have been linked to specific infections in neonatal units

• That the timings for Lucy’s involvement (as agreed upon by the prosecution and defense) might not add up

III. What do I think?

Is it possible that she committed a crime? Well, from what I’ve read I’m not sure there was enough concrete evidence to support that a crime was committed, but she was found guilty in a court of law. If you’re curious and want to go down the rabbit hole, follow this footnote.9

More importantly, though my opinion has been formed by what I’ve read: I can not know. I am neither a police nor a judge and I have not spent months or years poring over the original data and files. My opinion doesn’t matter in any significant way, precisely for those reasons. The sources of data which I have read may be flawed, biased or falsified. They may have omitted pertinent data. If I had to make a decision based on the data I’ve seen to date, I haven’t seen convincing support for either that a crime was committed or that Lucy was guilty. However, this could just be a consequence of not reviewing all the data in a systematic manner. There is much evidence that the prosecution presented which I haven’t presented in this article as I found it circumstantial and ambiguous. I’ve listed some of the main evidence in this footnote.10

There are several haunting similarities with the case of Lucia de Berk, a Dutch nurse that was convicted for several murders and attempted murders of patients. Just as in the two cases above, the main evidence was that she was present when patients died, there was one blood test (which turned out to be unreliable) and flawed statistical reasoning. Seven years after receiving her sentence, she was acquitted, when the court concluded that all deaths were in fact natural.11

The purpose of this text is not to defend Lucy. Regardless of her innocence or guilt, there are important lessons here. A court proceeding can be seen as a well-demarcated thought experiment, with lessons both for the legal proceedings, as well as decision-making in general. The lesson isn’t that internet sleuths are smarter than lawyers. Or that judicial systems are perfect or broken. It’s rather that reality is tricky, and there are specific areas where medicine is difficult for outsiders to interpret and decipher.

IV. Empathy for others

One last thing, before we summarize the meta-learnings. For one moment, just assume that Lucy Letby truly is entirely innocent, just like Lucia de Berk was.

Let that premise ruminate in your mind.

Imagine her perspective, being sentenced to many life sentences for something she didn’t do. The world cursing her name and seeing her as a symbol of evil. All her friends, family and colleagues distancing themselves from her, despite her innocence.

What are you feeling now? Do you feel frustration? Outrage about the court proceedings? An inability to understand how the jury didn’t pay attention to the statistical patterns and the flaws in the insulin test? Why the defense didn’t summon a single medical expert as a witness?

Hold on to the feelings and then erase the assumption of innocence.

Letby being innocent isn’t a conspiracy theory in any way. But I think that many people who believe in conspiracies experience the exact same emotions that you may be feeling now. Frustration over injustice and other people’s ignorance and disinterest. This case, beyond illustrating the complex world of medicine, offers a glimpse into how it can feel to believe a conspiracy theory. These glimpses are important - they let us understand and empathize with people who believe in conspiracy theories, and aid us in discussions with them.

Conspiracy theories are both emotionally and socially driven.12

• Emotional: Our attraction to conspiracy theories is driven by intuition. As a medical doctor, I can empathize with another clinician in this case and think "Surely another clinician couldn't be so evil?" I also feel that clinicians have an increasing lack of control of how healthcare works. "Couldn't this be the same thing, Lucy being thrown under the bus, with no control of the situation?”

• Social: Conspiracy theories are also a way to uphold a strong in-group identity and to protect something against a threatening outgroup. I feel belonging to the group of science and statistics nerds. Digging into the details of the case makes me feel (right or wrong) that I'm protecting science against less knowledgeable laymen.

So is the possibility of Letby being innocent simply fueled by cognitive biases? I don’t think so, because our conspirogenic biases can also explain a bias towards believing she is guilty.

• Sense-making: Our brains want to make sense of the world, by linking together stimuli (apophenia) and finding agency. We know that we have a bias to perceive events as caused by agents acting with intent, rather than simply accepting occurrences as random. Could random chance cause peaks and troughs of infant mortality? Statistics say yes, but our gut says no and wants to find some kind of agency involved in something so tragic and terrible.

• Injustice: Few things are more horrible and unjust than someone intentionally murdering infants. The idea itself fuels anger, which in turn can increase the risk of believing in a conspiracy, like an otherwise entirely normal nurse having a dark side and sneaking around killing infants.

Research on belief in conspiracy theories offers a glimpse into what nudges people into believing things that seem outright crazy. But more importantly, it should help us be kinder to those who believe in conspiracy theories. It should help us understand their perspective and aid us in talking to them and hopefully nuancing their views. It should make us humble about how easily we are affected by biases, especially when the subject matter and purported evidence is complex.

IV. Emmetropes

• It is difficult for non-expert laymen to understand medicine. This can result in overconfidence in…

  • Medical data (lab tests, x-ray findings), without being aware that there are always sources of error and limitations

  • How much clinicians understand of individual patient’s pathophysiology, and that patients often deteriorate without us knowing exactly why

  • Medical experts, without knowing that experts might disagree on how to interpret a certain datapoint, or finding

• It’s important to be epistemologically humble. Expect significant unknown unknowns when assessing a domain outside of one’s expertise, in order to minimize the risk of drawing faulty conclusions

  • Statistics is complex, and it’s unrealistic to expect laymen to have a clear understanding of statistics. Biases like the Texas Sharpshooter Fallacy, confirmation bias, base rate fallacy, and pre-decisional distortion distort our understanding and increase the risk of faulty decisions. Important decision-making processes (including trial proceedings) should actively be designed to minimize these biases, regardless of who is on trial and for what.

• Deep medical and scientific competence is essential in legal settings, in order to ensure a fair trial. This can in part, but not fully, be mitigated by involving an impartial third-party medical expert, as was the case in the ALB-case, which involved a governmental agency with expertise in assessing medical incidents and evidence

• Be aware of emotional, social, and sense-making biases, which increase the risk of falling for conspiracy theories

The first time I entered, I saw a completely ordinary emergency room: a gray-speckled, easy-to-clean vinyl floor, a CPR poster on the wall, and Post-it notes taped to the computer screen. Harsh fluorescent lamps casting an unflattering light. During a resuscitation call, the room seemed chaotic, but I soon realized that I was seeing a well-coordinated and intricate dance.

The dancers moved with rehearsed and synchronized steps, fighting to hold on to those whose time had not yet come. This was where healthcare was put to the test. There was something heroic and, above all, meaningful in what unfolded on stage: witnessing the entropy of maladies fought by a choreography of art, knowledge and science.

A year later, I had been away from the emergency department for around six months. I had worked on a project in primary care, immersing myself in financial penalty and bonus systems, educational programs and statistics, and learning about how many emergency visits are preventable. When I returned to the emergency room it had transformed.

The lighting had shifted to a more somber glow, with shadows in the corners. Suddenly the dance was no longer heroic. Many heart failure patients who entered should not have come at all, but should have received help earlier. The dance was now set to a lamentoso, a melancholic melody of all the patients who had fallen through the cracks. The important thing was not what happened in the room, but what hadn't happened the day, week, and month before the patient came in. The walls of the room collapsed like curtains falling off their rods. I saw how the illuminated dance stage was actually a small part of a much larger stage, a whole healthcare system that stretched far beyond the hospital walls.

A few years later, after completing a fellowship in medical technology, the room had transformed once again. I noticed something previously hidden in the scenery: the reason we could do so much for the patients was the technology in the room: devices, tubes, drugs. Outside the hospital, you can't examine or treat much, but in the emergency room all these miraculous tools made everything possible. Now, all the spotlights were shining directly on them.

I painfully realized that the dance in the emergency room depended of decades of work by engineers, researchers, and entrepreneurs on developing new technologies that improved healthcare. Suddenly, an almost endless stream of extras poured onto the stage: engineers and researchers frantically measuring and tinkering with all the technology in the room. When the emergency pager beeped, they rushed to leave. By the time the patient arrived, they were all gone, both in time and space, but they had left traces: lots of devices made of plastic, silicon and metal, and a refrigerator filled with drugs. And after everyone had left the room and the sensor-controlled lights had turned off, the room continued to hum with the quiet whirring of machines and monitors.

I visited the room one last time, many years later when the department was moving to a new hospital. It was emptied of staff, patients and equipment. For the first time, it echoed when I raised my voice. The silence and absence of background noise created an unnatural feeling, like a theater stage after the performance had ended and everyone had left. The magic was gone. For the first and last time, I saw the room for what it had always been: just a room.

Mammals often have two eyes because multiple perspectives are needed to see depth. However, to see and understand oneself, two eyes are not enough; you need a reflection. Like a magical mirror in a fairy tale, the room transformed itself and revealed things I otherwise would not have been able to see. And in turn, perhaps it transformed me as well, by showing me that healthcare had so many more sides than I knew.

My emergency department sadly no longer exists. It was replaced nearly a decade ago, a victim of time which it so valiantly fought to protect its patients from. But I hope there are other magical emergency rooms out there, so that others get the chance to see all the beautiful, important and complex sides of healthcare.1