<![CDATA[Biomic Insights]]>https://biomichealth.substack.comhttps://substackcdn.com/image/fetch/$s_!LRva!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe78ea433-25ee-4405-b082-ec574ed20db7_813x813.pngBiomic Insightshttps://biomichealth.substack.comSubstackFri, 04 Sep 2026 08:01:54 GMT<![CDATA[Could Your First Healthcare AI Job Be at Microsoft?]]>https://biomichealth.substack.com/p/could-your-first-healthcare-ai-jobhttps://biomichealth.substack.com/p/could-your-first-healthcare-ai-jobTue, 01 Sep 2026 15:15:33 GMT

Healthcare AI careers are evolving rapidly, and this week’s edition highlights one of the most intriguing opportunities currently available to clinicians: Clinical Specialist at Microsoft AI in London. As healthcare organisations move beyond pilot projects and begin integrating AI into everyday clinical workflows, there is growing demand for clinicians who can bridge the gap between technology and frontline care. This role offers a rare entry point into the field, focusing on AI implementation, workflow redesign, and real-world adoption rather than software engineering. Alongside this opportunity, we’ve curated the latest jobs, conferences, research, and expert perspectives shaping the future of healthcare AI.

🎧 Want to listen to this newsletter on-the-go? Check out our podcast summary below:

Jobs & Opportunities:

  • 💻 Clinical Specialist - Microsoft AI (London): Entry-level clinical-facing role supporting adoption of AI-enabled healthcare technologies. Ideal for clinicians interested in implementation, workflow redesign, and translating AI capabilities into real-world healthcare settings.

  • 📊 Healthcare Operations & EHR Systems Expert - Synthires (Remote): Contractor role providing healthcare domain knowledge to AI development teams working with electronic health records and operational healthcare data. Exposure to health informatics, AI evaluation and workflow optimisation.

  • 📝 Healthcare Expert Commentator - Authority.inc (Remote): Clinicians review healthcare content and provide expert feedback used to improve predictive and analytical AI systems. Low time commitment and suitable alongside medical training.

  • 🌍 AI and Health Researcher - Terra API (London): Early-stage opportunity within a digital health and AI company developing health-data platforms. The role provides exposure to AI research, healthcare datasets, product development, and translation of emerging technologies into clinical applications.

  • 🔬 Clinical Reviewer, AI Systems - Mercor (Remote): Clinical domain expert role focused on reviewing AI-generated medical content and documentation. Suitable for clinicians looking to gain experience at the intersection of clinical reasoning, evaluation science, and AI model development.

Upcoming Events:

Breaking News and Research:

Opinions:

Reading about the latest developments is one thing. Understanding how the technology actually works is another.

The Deep Dive is our premium newsletter that unpacks complex healthcare AI concepts, from machine learning and generative AI to model evaluation, safety, regulation, and real-world deployment. Perfect for clinicians looking to strengthen their technical understanding, excel in healthcare AI interviews, or become more effective leaders in digital health.

Check Out The Deep Dive

Biomic Insights is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

]]><![CDATA[Opportunities at TidalSense, Everlab, Heidi & Other Healthcare AI Startups]]>https://biomichealth.substack.com/p/opportunities-at-tidalsense-everlabhttps://biomichealth.substack.com/p/opportunities-at-tidalsense-everlabTue, 25 Aug 2026 16:34:16 GMT

This edition of Biomic Insights highlights several UK startup and health-tech roles spanning clinical research, product evaluation, analytics, clinical operations, and regulatory affairs - many of which do not require a technical background. Alongside these career opportunities, we feature upcoming conferences that offer exposure to healthcare AI, digital innovation, and health-tech entrepreneurship. We also review recent developments in clinical AI implementation, emerging research, and commentary on regulation, workforce implications, and the evolving role of clinicians in an increasingly technology-enabled healthcare system.

🎧 Want to listen to this newsletter on-the-go? Check out our podcast summary below:

Jobs & Opportunities:

  • 💡 Clinical Research Scientist - TidalSense (UK Health-Tech Startup): Early-career clinician-facing role supporting clinical evidence generation, regulatory activities, and scientific communication for AI-enabled respiratory diagnostics. Particularly relevant for doctors interested in medical devices, AI diagnostics, and digital health startups.

  • 📊 Medical Insights Analyst (ISR Recruitment, UK): Healthcare analytics role involving interpretation of medical and health-related data to support strategic decision-making. While not a pure AI position, it provides strong exposure to the data and analytics workflows that underpin healthcare AI products.

  • 🩺 Clinical Lead (UK) - Everlab (London, hybrid): Founding clinical leadership role at an AI-first preventative healthcare startup launching in the UK. The successful candidate will build Everlab’s UK clinical operations from the ground up, recruit and oversee a network of clinicians, shape clinical pathways and diagnostic protocols, and lead relationships with regulators including the CQC and MHRA.

  • 🚀 Expression of Interest Program - Heidi Health (UK): Heidi develops AI clinical scribe technology and maintains an ongoing talent pipeline for healthcare professionals interested in product, partnerships, clinical strategy, and implementation roles.

  • 🧪 Clinical Evaluation Specialist (£75k-£90k + Equity) - Jack & Jill (London, hybrid): Early-career clinical evaluation and regulatory role at a high-growth digital health startup developing medical software. The position focuses on reviewing clinical evidence, assessing safety and performance data, supporting regulatory submissions, and working closely with engineers and product teams to bring AI-enabled healthcare technologies to market.

Upcoming Events:

Breaking News and Research:

Opinions:

  • 📜 Regulating AI in healthcare: a moving target (The BMJ): This editorial discusses the UK’s evolving approach to healthcare AI regulation, arguing that safety oversight cannot stop at pre-market approval. The piece highlights concerns around model drift, automation bias and post-deployment monitoring—key issues for anyone interested in how AI will be governed within the NHS.

  • 🌍 Global advances in health artificial intelligence: a workforce imperative (The Lancet): Rather than focusing on model performance, this viewpoint examines how AI might address clinician shortages, administrative burden, burnout and workforce sustainability. Particularly relevant to NHS trainees, it frames AI as a tool for preserving clinical capacity and improving retention rather than replacing healthcare professionals.

  • 🤖 Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care? (JAMA): This JAMA Perspective explores one of the biggest questions in healthcare AI: should the future involve clinicians using AI, or AI acting more independently? The article weighs the advantages and limitations of physician-led versus autonomous AI-supported care and provides a good framework for thinking about the future role of doctors.

And that’s a wrap! 🎬 See you in the next edition.

Biomic Insights is a reader-supported publication. To receive new posts and support our work, consider becoming a free or premium subscriber.

Subscribe now

Leave a comment

]]><![CDATA[How machine learning actually works — no maths]]>https://biomichealth.substack.com/p/how-machine-learning-actually-workshttps://biomichealth.substack.com/p/how-machine-learning-actually-worksSat, 22 Aug 2026 18:03:17 GMT

🔓 This week’s Deep Dive: what’s happening inside the black box — loss functions, gradient descent, neural network layers, and why “the model learned it” is a sentence with real mechanics behind it. Plus the four questions this understanding lets you ask. [Read the Deep Dive] (premium)

🎯 In 5 minutes you’ll be able to:

  1. Describe how a model learns, using an analogy you can repeat in a meeting

  2. Explain overfitting — and why it’s the reason models flatter themselves

  3. Recognise the three things every model needs, and what happens when one is weak

You don’t need the mathematics. You do need the mechanism — because almost every failure we’ve covered so far (Weeks 1–5) is a mechanism failure wearing different clothes.

Here’s the whole thing, honestly, in one page.

The medical student analogy 🎓

Think about how a student learns to read chest X-rays.

They see a film. They guess: “pneumonia.” A consultant says: “no — that’s a mass.” The student adjusts something in their internal sense of what pneumonia looks like. Next film. Guess. Correction. Adjust. Repeat, four thousand times.

That’s it. That’s machine learning.

  • The films are training data.

  • The guess is a prediction.

  • The consultant’s correction is the label — the right answer.

  • “Adjust something internally” is the weights being updated.

  • Four thousand repetitions is training.

💡 KEY CONCEPT

A model doesn’t learn rules. It learns adjustments. Nobody tells it “pneumonia looks like consolidation.” It shifts millions of internal numbers until its guesses stop being corrected. Which is why nobody — including its creators — can fully state what it learned.

The three ingredients 🧂

Every model needs exactly three things. Weakness in any one caps everything.

1️⃣ Data — the examples. Enough of them, and representative of the world it’ll meet. 2️⃣ Labels — the right answers. As reliable as the consultant giving corrections. 3️⃣ Objective — a definition of “getting better,” so it knows which direction to adjust.

Sit with the second one, because it’s the one that bites in medicine. Our student learns from a consultant. If that consultant is sometimes wrong, the student learns those errors as truth — confidently, invisibly. And in clinical AI our “consultant” is often a radiology report, a diagnostic code, or a chart review: all imperfect, all systematically imperfect in particular directions.

A model can never be more right than the labels it learned from. It can only be more consistent. ⚠️

Overfitting: the student who memorised the exam 📚

Our student has a shortcut available. Instead of learning what pneumonia looks like, they could memorise: “film #47 is pneumonia, film #48 is normal.” Perfect performance on the practice set. Useless on a new patient.

That’s overfitting — and it’s the single most important concept in this article.

Models are excellent memorisers. Given enough capacity and not enough data, a model will happily memorise the noise, the quirks, the individual cases — and then report spectacular accuracy on the data it memorised.

This is why held-back test data exists: films the student has never seen, used once, at the end. It’s the only honest exam. And it’s why Week 3’s evidence ladder is built the way it is: every rung is a harder exam, further from what the model memorised.

💡 KEY CONCEPT

The gap between training performance and real-world performance is memorisation, not intelligence. When a published number looks too good, “how separate was the test data, really?” is nearly always the right question.

Why models take shortcuts 🚩

Our student has another failure available: noticing that pneumonia films at their hospital tend to come from the portable machine — and learning the machine, not the disease.

They’d score brilliantly. They’d have learned nothing about pneumonia.

Models do this constantly, because they’re optimising for “stop being corrected,” not “understand the disease.” Any signal that predicts the label will do. This is shortcut learning (Week 1), and it’s not a bug — it’s the objective working exactly as specified. The model was never asked to understand. It was asked to be right. 🎯

What this buys you 🧭

Understanding this mechanism upgrades every question you ask:

  • “What did it learn from?” → because it can’t beat its labels

  • “How separate was the test set?” → because memorisation flatters

  • “What might it be keying on instead?” → because shortcuts are the default, not the exception

  • “Does the objective match what we actually want?” → because the model optimises the objective, not your intention

📌 Remember three things

  1. Guess → correction → adjust, repeated. That’s all learning is.

  2. A model can’t be more right than its labels — only more consistent.

  3. Overfitting is memorisation. Held-back data is the only honest exam.

🔓 The Deep Dive opens the black box: • Loss functions — how “getting better” is actually defined, and why the definition determines everything • Gradient descent — the “adjust something” step, made concrete (still no maths required) • What layers do — why deep networks build features from edges to anatomy • Why nobody can explain a prediction — the honest version of the interpretability problem, and what saliency maps do and don’t tell you • The four mechanism questions to ask any vendor — and what good answers sound like [Read the Deep Dive] (premium)

The Healthcare AI Field Guide is published by Biomic Health. Subscribe:

Biomic Insights
🏥🤖 The all in one healthcare AI newsletter. Curated opportunities, news, research, perspectives, resources and events. Delivered each Sunday 📩. Plus articles, deep dives and interviews.
By Biomic Health

Website: https://www.biomichealth.com

Free resource hub: https://resources.biomichealth.com

]]><![CDATA[Inside the black box]]>https://biomichealth.substack.com/p/inside-the-black-boxhttps://biomichealth.substack.com/p/inside-the-black-boxSat, 22 Aug 2026 17:55:16 GMT🔒 Premium Deep Dive. The Foundation gave you the mechanism in analogy. This is the mechanism itself — loss, gradients, layers, and the honest version of why nobody can explain a prediction.

🎯 In this Deep Dive you’ll learn:

  1. What a loss function is — and why choosing it is a clinical decision disguised as a technical one

  2. How gradient descent works — the “adjust something” step, concretely

  3. What layers actually do — from edges to anatomy

  4. Why explanation is genuinely hard — what saliency maps do and don’t show

  5. The four mechanism questions — and what good answers sound like

Read more

]]><![CDATA[The Latest Opportunities: From Responsible AI Specialist and Clinical Product Specialist Roles to AI Training and Safety Research]]>https://biomichealth.substack.com/p/the-latest-opportunities-from-responsiblehttps://biomichealth.substack.com/p/the-latest-opportunities-from-responsibleTue, 18 Aug 2026 16:01:57 GMT

This week’s edition of Biomic Insights highlights five roles spanning AI governance, clinical product development, model training and AI safety research, each offering a different route into the sector for medical students, resident doctors and practising clinicians. We also feature selected conferences, recent healthcare AI developments and noteworthy research and commentary from leading journals. Together, these provide a snapshot of a rapidly evolving field that is creating new professional pathways while reshaping how healthcare technologies are developed, evaluated and implemented in practice.

🎧 Want to listen to this newsletter on-the-go? Check out our podcast summary below:

Jobs & Opportunities:

Upcoming Events:

Breaking News and Research:

  • 🧩 Google DeepMind’s AMIE expands into real-time video consultations: Google Research and Google DeepMind reported that AMIE, their medical AI research system, can now conduct real-time clinical video consultations using visual and auditory information as well as conversational history. In a randomized study with simulated patients and primary care physicians, evaluators rated AMIE favorably across history-taking, diagnostic accuracy, management planning and communication. Although still a research system, it offers a glimpse of how future AI tools could support remote consultations and triage.

  • 🩺 Google launches “Health Guardian” AI monitoring features for wearables: Google announced new Health Guardian capabilities for Pixel Watch and Fitbit devices that continuously monitor physiological signals and generate AI-supported health insights. The system aims to detect trends in areas such as insulin resistance, blood pressure, sleep breathing quality and respiratory events, with monthly summaries and coaching delivered through the Google Health app. For clinicians, this signals continued movement toward AI-assisted preventive and longitudinal health monitoring.

  • 🏥 Singapore hospital group saves 31,000 administrative hours using AI: Healthcare IT News highlighted how IHH Singapore is using AI to anticipate demand surges and proactively adjust staffing and patient flow. The organization reported saving approximately 31,000 administrative hours annually. This is another example of healthcare AI moving beyond diagnostics into operational management and workforce planning.

  • 🤖 Clinicians vs. artificial intelligence in patient outcome prediction in the intensive care unit (npj Digital Medicine): This study directly compares clinician judgement with AI models for predicting ICU outcomes. For trainees in acute medicine, anaesthetics or critical care, it’s a useful reality check on where AI currently outperforms humans, where it does not, and what “human-AI collaboration” might actually look like at the bedside rather than in theory.

  • 🩻 Histological aging signatures for monitoring tissue-specific aging and disease (Nature Medicine): Researchers used whole-slide histopathology images from 40 tissue types to derive AI-based “aging clocks” linked to tissue integrity and disease. For pathologists, oncologists and clinician-scientists, this is a good example of AI moving beyond diagnosis toward extracting entirely new biological signals from routine clinical data.

Opinions:

And that’s a wrap! 🎬 See you in the next edition.

Biomic Insights is a reader-supported publication. To receive new posts and support our work, consider becoming a free or premium subscriber.

Subscribe now

Leave a comment

]]><![CDATA[Why so much clinical AI solves the wrong problem]]>https://biomichealth.substack.com/p/why-so-much-clinical-ai-solves-thehttps://biomichealth.substack.com/p/why-so-much-clinical-ai-solves-theWed, 12 Aug 2026 19:37:36 GMT

🔓 This week’s Deep Dive: the economics behind the mismatch — why labelled data bends the whole field, the value equation that predicts whether a tool will matter, the “so what” ladder, and a one-page problem-selection canvas to write before you look at a single vendor. [Read the Deep Dive] (premium)

🎯 In 5 minutes you’ll be able to:

  1. Explain why clinical AI clusters around certain problems and ignores others

  2. Spot the three mismatches that make a technically excellent tool clinically pointless

  3. Apply three questions that test whether a problem is worth solving at all

Ask a clinician where their week actually goes, and you’ll hear about documentation, chasing results, discharge letters, referrals, coordinating with other teams, and the slow grind of finding information that already exists somewhere in the record.

Now look at where clinical AI development goes. Overwhelmingly: diagnostic pattern recognition. Detecting the nodule, flagging the bleed, grading the retina.

Those aren’t the same list. That gap is this week’s subject — and it isn’t because engineers are foolish or clinicians are unclear. It’s structural.

Why the gap exists 🧲

Machine learning needs labelled examples. So the field flows toward problems where labels already exist as a by-product of care:

  • A chest X-ray has a report attached. That’s a label. ✅

  • A pathology slide has a diagnosis. That’s a label. ✅

  • A structured EHR field has a value. That’s a label. ✅

  • The reason a discharge letter took nine days, the corridor conversation that resolved a referral, the judgement call that avoided an admission? No label. ❌

💡 KEY CONCEPT

Data gravity, not clinical need, selects the problems. The field’s attention follows where labelled data is cheapest to obtain — which correlates poorly with where clinical burden is heaviest. This single fact explains most of the mismatch between what gets built and what would actually help.

Publication and regulation reinforce it. A tidy imaging classification task produces a clean AUROC and a publishable paper; “we made discharge letters 20% faster” is harder to measure and less glamorous. And a narrow diagnostic claim is an easier regulatory path than a tool that touches a whole messy workflow.

The three mismatches ⚠️

When a tool solves the wrong problem, it’s usually one of these three.

1️⃣ Burden mismatch — solving the small problem, brilliantly. The tool works, and the thing it fixes was never what cost you time. A model that shaves seconds off an interpretation you found easy, while the two hours of documentation stays untouched. Test: if this worked perfectly, how much of my week changes?

2️⃣ Decision mismatch — answering a question nobody was struggling with. The most seductive version. Tools tend to perform best on obvious findings and worst on subtle ones — because obvious findings are well represented in training data and easy to label. So you get confident confirmation of the fracture you’d already seen, and silence on the one you’d have missed. The tool is most reliable exactly where you needed it least. Test: does it help on the cases I find hard, or only the ones I find easy?

3️⃣ Value mismatch — the accuracy that changes nothing. The performance improvement is real and no decision moves because of it. If a risk score goes from 0.79 to 0.83 AUROC but every patient still gets the same pathway, you’ve bought a better number, not better care. Test: which decision changes, for whom, and what happens differently as a result?

💡 KEY CONCEPT

Accuracy is a property of the model. Value is a property of the decision it changes. No decision change → no value, no matter how good the metric. This is why “it’s 94% accurate” is an answer to a question you didn’t ask.

The three questions 🧭

Before evaluating any tool — before the Four Gates, even — test the problem:

❓ Would it change a decision? Name the decision, the person making it, and the moment it’s made. If you can’t, there’s nothing to buy yet.

❓ Would anyone notice if it vanished? Imagine it running for six months and then switched off silently. Who complains? If nobody, it wasn’t solving a felt problem.

❓ Who benefits, and who does the extra work? Value that accrues to the organisation while the effort falls on already-stretched clinicians is the most reliable predictor of quiet abandonment. Misaligned effort kills more deployments than bad models do.

The reframe 🔄

The useful move isn’t to demand AI for admin instead of imaging — plenty of diagnostic AI is genuinely valuable, and the imaging backlog is real. It’s to start from the burden and work outward, rather than starting from the technology and looking for somewhere to put it.

Ask your team: what are the three most expensive, most repeated, most morale-draining tasks here? Then ask which of those a machine could plausibly touch. That list — not a vendor’s product catalogue — is your actual AI strategy.

📌 Remember three things

  1. Data gravity picks the problems. What gets built follows where labels are cheap, not where the pain is.

  2. Tools are often most reliable where you needed them least — obvious findings, easy cases.

  3. Value lives in the decision, not the metric. No decision changes → no value, whatever the accuracy.

🔓 The Deep Dive goes into the economics:

• Why labelled data bends the field — the cost structure of labels, and the publication and regulatory incentives that compound it

• The value equation: decision change × frequency × consequence − (workflow cost + failure cost), worked through with examples

• The "so what" ladder — five rungs from model metric to system outcome, and why most tools stop at rung one

• Where the unglamorous value actually is — the problem classes that are underserved precisely because they're hard to label

• The problem-selection canvas: a one-page document to write before you meet a vendor

[Read the Deep Dive] (premium)

The Healthcare AI Field Guide is published by Biomic Health. Subscribe:

Biomic Insights
🏥🤖 The all in one healthcare AI newsletter. Curated opportunities, news, research, perspectives, resources and events. Delivered each Sunday 📩. Plus articles, deep dives and interviews.
By Biomic Health

Website:

https://www.biomichealth.com

Free resource hub:

https://resources.biomichealth.com

]]><![CDATA[The economics of the wrong problem]]>https://biomichealth.substack.com/p/the-economics-of-the-wrong-problemhttps://biomichealth.substack.com/p/the-economics-of-the-wrong-problemWed, 12 Aug 2026 19:33:15 GMT🔒 Premium Deep Dive. The Foundation named the mismatch. This is why it exists economically, how to predict whether a tool will matter before you buy it, and the one-page canvas to write before you meet a single vendor.

🎯 In this Deep Dive you’ll learn:

  1. Why labels bend the field — the cost structure that decides which problems get solved

  2. The value equation — a back-of-envelope model that predicts whether a tool will matter

  3. The “so what” ladder — five rungs, and why most evidence stops on rung one

  4. Where the underserved value is — problem classes ignored precisely because they’re hard

  5. The problem-selection canvas — your pre-vendor one-pager

1. The label economy 🧾

Every supervised model needs paired examples: input, and the right answer. Where that pairing already exists as exhaust from normal care, a product is cheap to build. Where it doesn’t, someone must manufacture it — and that changes everything.

Free labels (by-products of care): radiology reports attached to images, pathology diagnoses attached to slides, coded outcomes, lab values, discharge codes. Millions of examples, no extra cost, retrospectively available.

Read more

]]><![CDATA[Flexible AI-training roles, clinical engineering opportunities and the latest evidence shaping safe adoption of clinical AI]]>https://biomichealth.substack.com/p/flexible-ai-training-roles-clinicalhttps://biomichealth.substack.com/p/flexible-ai-training-roles-clinicalTue, 11 Aug 2026 22:31:32 GMT

This edition explores how clinicians can begin contributing to healthcare AI through flexible model-evaluation work, scientific review and hybrid clinical-engineering roles. The opportunities span remote freelance projects that fit alongside NHS training and full-time positions building tools for frontline care. Beyond careers, we highlight five major UK events focused on implementation, policy and evidence, alongside developments in diagnostic AI, point-of-care clinical support and AI-designed therapeutics. New research and commentary also examine automated imaging, LLM-generated evidence summaries, surgical video analysis and responsible academic use, showing why clinical judgement, validation and governance remain central as adoption accelerates across healthcare systems worldwide today.

🎧 Want to listen to this newsletter on-the-go? Check out our podcast summary below:

Jobs & Opportunities:

  • 🩺 Remote Clinical AI Evaluator - entertheloop (UK-wide, remote, flexible/paid): Clinicians evaluate AI-generated answers against NICE, BNF, Royal College and other UK standards, scoring them for safety and clinical accuracy. Open to GMC-registered doctors from SHO level upwards, with work allocated according to experience, making it particularly compatible with resident training.

  • 🧠 Doctors - AI Training - Prolific (UK, remote, freelance): Flexible project-based work reviewing clinical scenarios and evaluating the accuracy, safety and reasoning of advanced AI models. Advertised remuneration is approximately $80-$150 per completed AI task, offering qualified doctors direct experience in clinical model evaluation without leaving NHS employment.

  • 🔬 Medical/Biological Researcher - AI Trainer - 10x.Team (UK/EU, remote, freelance): Review AI-generated scientific material, develop realistic research scenarios and identify methodological or factual errors. The position pays €30-€40 per hour for approximately 8–20 hours per week and could suit medical students with substantial research experience as well as academically active resident doctors.

  • 🧪 Doctors and Medical Experts for AI Research Projects - Aptura (UK, remote, contract/rolling recruitment): Contribute to healthcare-AI projects through clinical data annotation, model evaluation, prompt development and the creation of clinical scoring rubrics. This is a flexible route for doctors seeking initial experience in human-in-the-loop evaluation and AI safety work alongside clinical training.

  • 💻 Clinical Engineer - Eolas Medical (London or Belfast, full-time): Hybrid clinical and software-development position building AI-powered medical knowledge tools used by frontline clinicians. Eolas explicitly welcomes doctors with minimal previous programming experience as well as established clinician-engineers; the advertised salary is £60,000-£110,000 plus stock options.

Upcoming Events:

Breaking News and Research:

  • 🩻 Twelve US health systems establish a Diagnostic AI Consortium: Twelve health systems, including University Hospitals, Stanford Health Care and Mayo Clinic, have joined Aidoc to evaluate how diagnostic AI affects safety, diagnostic speed and quality of care. Crucially, the consortium plans to share evidence and governance lessons beyond its participating hospitals - potentially offering a useful model for NHS organisations deciding how imaging and clinical AI should be validated, monitored and scaled.

  • 💊 UpToDate expands its clinical AI tools to approximately 2,500 US hospitals: Wolters Kluwer reports that UpToDate Expert AI is now available across roughly 2,500 hospitals and health systems, alongside a new drug-dosing capability tested using real-world medication questions. Its rapid adoption shows how generative interfaces are moving into mainstream point-of-care resources, making source transparency, evidence currency and clinician verification increasingly important skills for doctors.

  • 🩻 Automated breast ultrasound meets artificial intelligence: a new accurate classification tool for BI-RADS 3-4 breast lesions (Journal of Ultrasound): This original study evaluates whether AI applied to automated breast ultrasound can help distinguish lower-risk BI-RADS 3 lesions from suspicious BI-RADS 4 abnormalities. It is a clinically tangible example of AI supporting radiological classification and potentially reducing unnecessary biopsies - but, as with any diagnostic model, its usefulness will ultimately depend on external and prospective validation in representative screening populations.

  • 📚 Trial Files: Leveraging large language models to summarise practice-changing clinical trials for clinicians (PLOS One): Using information extracted from 568 randomised controlled trials, this study examines whether LLMs can generate concise and accurate summaries of research likely to change clinical practice. The results suggest that carefully designed systems could help doctors keep pace with new evidence, while highlighting why traceable references, structured quality assurance and clinician review remain essential before AI-generated summaries are trusted for clinical decisions.

  • 🧬 Generative design of bacteriophages with genome language models (Science): Researchers used genome language models to generate complete bacteriophage genomes, synthesising and experimentally testing hundreds of designs and identifying viable phages capable of infecting bacteria. Although this is early-stage biological engineering rather than immediate clinical practice, it points towards future AI-designed phage therapies for antimicrobial-resistant infections while raising substantial biosafety, governance and containment questions.

Opinions:

  • ✍️ Updated guidance for author use of AI in medical publication (JAMA): This editorial updates the JAMA Network’s rules on using AI in research and manuscript preparation, stressing disclosure, human accountability, protection of confidential information and verification of AI-generated material. It is particularly relevant if you use generative AI for literature reviews, abstracts or academic writing: the technology may assist authors, but responsibility for accuracy, originality, citations and research integrity remains entirely human.

  • 🎥 Surgical scene understanding and the emerging challenge of clinical translation (npj Digital Medicine): This surgical-AI commentary argues that recognising instruments or operative phases on video is not enough to demonstrate meaningful clinical value. The authors call for models to be connected to patient and team outcomes, supported by longitudinal multimodal datasets and evaluated through implementation research. It is highly relevant to surgeons interested in computer vision because it shifts the question from “Can the algorithm identify the scene?” to “Does that information actually improve an operation?”

  • 🧪 AI-based augmentation of oncology clinical trials (Nature Reviews Clinical Oncology): This review argues that AI’s most credible immediate contribution to cancer trials lies in supporting patient identification, eligibility screening, data extraction and monitoring under human supervision. More ambitious applications - including synthetic control arms, digital twins and simulated trial outcomes - remain insufficiently validated. It offers a useful framework for separating deployable workflow improvements from claims that AI can replace conventional evidence generation.

And that’s a wrap! 🎬 See you in the next edition.

Biomic Insights is a reader-supported publication. To receive new posts and support our work, consider becoming a free or premium subscriber.

Subscribe now

Leave a comment

]]><![CDATA[Before your department says yes to an AI tool]]>https://biomichealth.substack.com/p/before-your-department-says-yes-tohttps://biomichealth.substack.com/p/before-your-department-says-yes-toSat, 01 Aug 2026 07:56:19 GMT

🔓 This week’s Deep Dive: the full evaluation playbook — the evidence ladder for AI studies, the metrics vendors don’t lead with, how to run a silent trial, the red-flag list, and a printable scorecard for your next procurement meeting. [Read the Deep Dive] (premium)

Read more

]]><![CDATA[The evaluation playbook]]>https://biomichealth.substack.com/p/the-evaluation-playbookhttps://biomichealth.substack.com/p/the-evaluation-playbookSat, 01 Aug 2026 07:50:00 GMT

🔒 Premium Deep Dive. The Foundation gave you the Four Gates. This is the working playbook behind them — the evidence ladder, the metrics, the silent trial, the red flags, and a scorecard you can take into a procurement meeting.

🎯 In this Deep Dive you’ll learn:

  1. The evidence ladder — how to rank any AI study in seconds, and the reporting checklists that audit a vendor’s paper for you

  2. The metrics beyond accuracy — calibration, subgroup performance, and drift, with the exact phrases for the meeting

  3. How a silent trial works — evaluating a tool on your own patients with zero clinical risk

  4. The ten red flags — vendor behaviours that predict failed deployments

  5. The scorecard — the Four Gates as a working document

Read more

]]><![CDATA[From Founder's Associate to AI Engineer: Breaking into Healthcare AI Startups]]>https://biomichealth.substack.com/p/from-founders-associate-to-ai-engineerhttps://biomichealth.substack.com/p/from-founders-associate-to-ai-engineerTue, 28 Jul 2026 11:29:56 GMT

This edition highlights the breadth of opportunities emerging within the UK healthcare AI startup ecosystem. From founder-level operational roles to engineering positions developing autonomous AI agents, these vacancies demonstrate the diverse routes into health technology for both clinicians and technical professionals. Alongside these opportunities, w…

Read more

]]><![CDATA[NHS Informatics and Endoscopic Imaging: Implementing Innovative Technology as a Clinical/Research Fellow]]>https://biomichealth.substack.com/p/nhs-informatics-and-endoscopic-imaginghttps://biomichealth.substack.com/p/nhs-informatics-and-endoscopic-imagingFri, 17 Jul 2026 14:19:14 GMT

For clinicians seeking a practical route into healthcare AI, this edition foregrounds two distinctive fellowships. The London Research Fellow role offers hands-on experience developing and validating AI for advanced endoscopic imaging, working across clinical, engineering and computer-science teams. The Mid Yorkshire Clinical Fellow post provides a diff…

Read more

]]><![CDATA[You're already using AI. You just don't call it that. 🩻]]>https://biomichealth.substack.com/p/youre-already-using-ai-you-just-donthttps://biomichealth.substack.com/p/youre-already-using-ai-you-just-dontSun, 12 Jul 2026 16:13:50 GMT Read more

]]><![CDATA[Inside the AI nobody consented to 🔍]]>https://biomichealth.substack.com/p/inside-the-ai-nobody-consented-tohttps://biomichealth.substack.com/p/inside-the-ai-nobody-consented-toSun, 12 Jul 2026 16:10:44 GMT🎯 In this Deep Dive you’ll learn:

  1. How deep-learning reconstruction works — and its two named failure modes (feature loss, hallucinated texture)

  2. What the ECG automation-bias evidence shows — the cleanest proof that machine accuracy ≠ system accuracy

  3. What alert fatigue predicts for every new AI alert entering your hospital

  4. Why triage AI redistributes risk without generating a single incident report

  5. How to run the ten-minute invisible AI audit — four conversations that produce your department’s first AI asset register

Read more

]]><![CDATA[MHRA Careers for Doctors: Medical Assessor and Innovation Expert Roles in Medicines and Healthcare Technology]]>https://biomichealth.substack.com/p/mhra-careers-for-doctors-medicalhttps://biomichealth.substack.com/p/mhra-careers-for-doctors-medicalWed, 08 Jul 2026 00:17:12 GMT

This edition of Biomic Insights focuses on two roles at the Medicines and Healthcare products Regulatory Agency (MHRA) that illustrate how doctors can contribute to the assessment and governance of emerging health technologies. The Medical Assessor post centres on evaluating clinical evidence submitted for medicines and determining their benefit-risk ba…

Read more

]]><![CDATA[How clinical AI learns — and the four ways it fails]]>https://biomichealth.substack.com/p/how-clinical-ai-learns-and-the-fourhttps://biomichealth.substack.com/p/how-clinical-ai-learns-and-the-fourTue, 30 Jun 2026 20:34:33 GMT

In the Foundation we sorted the vocabulary: AI contains machine learning, which contains deep learning, which contains LLMs — and almost every clinical tool is doing one of four jobs (perception, prediction, language, optimisation).

That’s enough to place a tool. It’s not enough to trust one. For that you need a feel for how these models learn, and — more importantly — the specific, repeatable ways they fall over in the messy reality of clinical work. None of this requires maths. It requires knowing where to point your scepticism.

How a model actually learns

Machine learning has one core ambition: generalisation. The model should perform well on patients it has never seen, not just memorise the ones it has.

To get there, you split your data three ways:

  • Training set — the examples the model learns from.

  • Validation set — used to tune the model as you build it.

  • Test set — held back, untouched, used once at the end to estimate real-world performance.

Read more

]]><![CDATA[What we actually mean when we say "AI"]]>https://biomichealth.substack.com/p/what-we-actually-mean-when-we-sayhttps://biomichealth.substack.com/p/what-we-actually-mean-when-we-sayTue, 30 Jun 2026 20:31:07 GMT Read more

]]><![CDATA[Exploring Heidi Health’s Clinical Associate Role and Newcastle University’s Digital Health Research Associate Opportunity]]>https://biomichealth.substack.com/p/exploring-heidi-healths-clinicalhttps://biomichealth.substack.com/p/exploring-heidi-healths-clinicalFri, 26 Jun 2026 15:44:37 GMT

This week’s edition examines two complementary routes into healthcare AI: clinical implementation and academic research. Heidi Health’s Clinical Associate role offers clinicians direct experience supporting the adoption of ambient AI, improving documentation workflows and translating frontline feedback into product development. Newcastle University’s Re…

Read more

]]><![CDATA[High-Risk AI under the EU AI Act - Why “used in healthcare” is not the whole test]]>https://biomichealth.substack.com/p/high-risk-ai-under-the-eu-ai-acthttps://biomichealth.substack.com/p/high-risk-ai-under-the-eu-ai-actSun, 21 Jun 2026 20:38:32 GMT

The Healthcare AI Field Guide

👩‍⚕️ Who This Is For

Medical students and resident doctors who want to understand how the EU AI Act applies to clinical technologies.

This topic is particularly valuable for anyone:

  • Evaluating healthcare AI companies

  • Working in digital health or clinical safety

  • Participating in technology procurement

  • Preparing for healthcare AI interviews

  • Considering a role in product development or regulation

🧩 The Core Idea

The EU AI Act uses a risk-based regulatory framework. AI systems are treated differently depending on the risks associated with their intended use. Broadly, systems may fall into categories including:

  • Prohibited AI practices

  • High-risk AI systems

  • Systems with specific transparency obligations

  • Systems presenting minimal or limited risk

Many clinically significant AI medical devices are likely to be classified as high-risk. However, the system is not classified as high-risk simply because it is used somewhere inside a hospital. For an AI system associated with regulated products such as medical devices, two important questions are:

  • Is the AI system itself a regulated product or a safety component of one?

  • Does the product require a third-party conformity assessment before it can be placed on the market?

Where both conditions are met, the AI system will generally fall within the EU AI Act’s high-risk framework.

🏥 A Clinical Analogy

Imagine three AI tools being used within the same hospital.

System 1: Administrative transcription tool

The system converts a clinician’s dictated letter into text. The clinician reviews and approves the letter before it is sent. Although errors could create problems, the system is not intended to diagnose disease or determine treatment.

System 2: Diagnostic imaging algorithm

The system analyses CT pulmonary angiograms and highlights suspected pulmonary emboli for clinical review. Its output may influence whether a serious diagnosis is identified and whether anticoagulation is started.

System 3: Autonomous deterioration alert

The system analyses observations and laboratory results and determines which patients require urgent escalation. A missed alert could delay treatment, while an incorrect alert could lead to unnecessary intervention. All three systems use AI. But they do not necessarily carry the same regulatory classification. The intended purpose and potential clinical consequences matter more than the fact that they all operate in a healthcare environment.

⚠️ Why This Happens in Healthcare

Healthcare organisations use AI for a wide range of activities:

  • Appointment scheduling

  • Clinical documentation

  • Diagnostic imaging

  • Pathology interpretation

  • Risk prediction

  • Treatment recommendations

  • Patient prioritisation

  • Workforce management

These systems create very different levels of risk. An AI tool that reorganises a clinic schedule is not equivalent to one that recommends whether a patient should receive chemotherapy. Regulation therefore focuses heavily on intended purpose.

This means asking:

  • What has the manufacturer designed the system to do?

  • What claims are being made about it?

  • Who uses the output?

  • What decision follows the output?

  • What could happen if the system is wrong?

🚨 Why This Matters

High-risk classification brings substantial regulatory responsibilities. Providers of high-risk AI systems may need to establish:

  • A risk-management system

  • Appropriate data-governance processes

  • Technical documentation

  • Automatic logging and record keeping

  • Clear instructions for users

  • Effective human-oversight measures

  • Appropriate accuracy, robustness and cybersecurity

  • Quality-management processes

  • Post-market monitoring

  • Procedures for reporting serious incidents

These are not simple administrative exercises. They affect how the system is:

  • Designed

  • Validated

  • Documented

  • Deployed

  • Monitored

  • Updated

A company that misunderstands its classification could build the wrong evidence and governance strategy from the beginning. A hospital that misunderstands the classification could purchase a product without asking the right safety and compliance questions.

🧠 What Most People Get Wrong

Many people assume:

“If a product has a CE mark, it has satisfied every relevant requirement.”

That is too simplistic.

An AI medical device may sit within overlapping regulatory frameworks.

For example, it may need to comply with:

  • The Medical Devices Regulation or In Vitro Diagnostic Medical Devices Regulation

  • The EU AI Act

  • Data-protection law

  • Cybersecurity requirements

  • National healthcare governance arrangements

The Medical Devices Regulation focuses on the safety and performance of the medical device. The EU AI Act adds requirements aimed specifically at risks created by AI, such as:

  • Data quality

  • Bias

  • Transparency

  • Human oversight

  • Logging

  • Ongoing AI-specific risk management

The two frameworks are intended to interact rather than operate as completely separate approval pathways.

🔄 It Is Not Just the Manufacturer’s Responsibility

Another common mistake is to assume that regulation concerns only the company building the model. The EU AI Act distinguishes between different actors, including:

Provider: The organisation that develops an AI system or has it developed and places it on the market under its own name.

Deployer: The organisation using the AI system under its authority. A hospital using a high-risk AI system may therefore have responsibilities of its own.

These can include:

  • Following the provider’s instructions

  • Assigning appropriately trained people to oversee the system

  • Ensuring input data are relevant and sufficiently representative where the hospital controls those data

  • Monitoring system operation

  • Retaining relevant logs

  • Responding appropriately when risks or serious incidents are identified

The precise obligations depend on the system and the circumstances in which it is used.

👁️ Human Oversight Does Not Mean “A Doctor Is Somewhere in the Loop”

The phrase “human in the loop” is frequently used in healthcare AI. However, meaningful human oversight requires more than placing a clinician near the technology. The person overseeing the system must be able to:

  • Understand its capabilities and limitations

  • Recognise automation bias

  • Interpret the output in context

  • Identify abnormal or unsafe behaviour

  • Disregard or override the output when appropriate

  • Escalate concerns

A clinician who automatically accepts every recommendation is not providing effective oversight. Equally, a clinician cannot meaningfully supervise a system if its output is incomprehensible, poorly presented or delivered without adequate training.

🧑‍⚕️ The Clinician’s Advantage

Clinicians are well placed to assess the real risk associated with an AI system. Technical teams may understand how a model was built. Regulatory teams may understand the legal framework. Clinicians understand what happens next.

They can identify:

  • Which clinical decision follows the output

  • Whether the output is genuinely actionable

  • How quickly someone must respond

  • What harm could result from an incorrect prediction

  • Whether the proposed human oversight is realistic

  • Whether users may become over-reliant on the system

This clinical insight is central to risk classification, validation and safe deployment.

🧠 What to Say in Interviews

“Under the EU AI Act, a system is not classified as high-risk simply because it is used in healthcare. Classification depends on its intended purpose, its relationship with regulated products such as medical devices and whether third-party conformity assessment is required.”

For an even stronger answer:

“Many clinically significant AI medical devices will be high-risk, but classification is only the starting point. Safe deployment also requires risk management, appropriate data governance, meaningful human oversight, monitoring and clear responsibilities for both providers and healthcare organisations.”

🚀 Practical Takeaways

When evaluating a healthcare AI product, ask:

  • What is its formally defined intended purpose?

  • Is it classified as medical-device software?

  • What medical-device risk class has been assigned?

  • Was a notified body involved in conformity assessment?

  • Is the system considered high-risk under the EU AI Act?

  • What evidence supports its clinical safety and performance?

  • How is human oversight supposed to work?

  • Who monitors the system after deployment?

  • What happens when its performance deteriorates?

  • Who reports a serious incident?

Do not accept “AI compliant” or “EU approved” as sufficient answers. Ask which legislation applies, what conformity route was followed and what evidence supports the claim.

🧠 30-Minute Exercise with Model Answers

Scenario

A company develops an AI system that analyses chest radiographs and identifies patients with a suspected pneumothorax.

The system:

  • Prioritises suspicious examinations on the radiology worklist

  • Produces a visible alert

  • Does not independently issue a final diagnosis

  • Is intended to accelerate radiologist review

  • Will be sold to hospitals throughout the European Union

1️⃣ Is this automatically high-risk because it is used in healthcare?

Answer

Read more

]]><![CDATA[Google's AMIE is leaving simulated consultations behind, but real-world research will determine whether conversational AI genuinely improves patient care]]>https://biomichealth.substack.com/p/googles-amie-is-leaving-simulatedhttps://biomichealth.substack.com/p/googles-amie-is-leaving-simulatedSun, 21 Jun 2026 13:06:59 GMT

This week’s Biomic Insights highlights five entry-level roles for clinicians interested in implementation, customer success and healthtech operations, alongside upcoming UK conferences focused on safe deployment. We also examine Google’s move to test conversational AI in real clinical settings, new evidence on emergency-department summarisation, and eff…

Read more

]]>