Another thing from that MHRA session I mentioned is the thinking around mistakes. What I took from Lawrence Tallon’s comments is that there are two elements: safety and then liability. Safety For drugs, there’s a chunky barrier to market because once you swallow it, you’re committed. The difference for AI is that it’s...

I tuned into the MHRA's "National AI Commission: Ask me Anything" session. It wasn’t a real AMA, but more of an update on the engagement work they've been doing on AI and healthcare in the UK. I think it was Henrietta Hughes who suggested reviewing AI output with your doctor. I like the idea. Is there enough time in a...

I'm working on a machine learning cancer classification problem, but we have only a handful of positive cases in the data. Thankfully. But that fact does make my job harder. It causes three main problems: We may not have enough variety to be able to find the true patterns in the data. This is a fundamental blocker, and...

The Guardian report on Biobank data being leaked onto GitHub is either no big deal or a massive screw-up, depending on your point of view. Either way, what I do is have a global gitignore file naming folders for confidential information. I put anything sensitive in there. If I want to push to Git, I have to try really,...

Reading: Are AI Tools Ready to Answer Patients’ Questions About Their Medical Care?, JAMA, 6 March 2026. OpenAI says that ChatGPT Health is built “to guide patients to health care professionals for diagnosis and treatment.” Unfortunately it currently under-triages emergency conditions and over-triages non-emergency...

Reading: If a therapy bot walks like a duck and talks like a duck then it is a medically regulated duck, npj Digital Medicine, 5 December 2025. Medical device regulation depends on the claims and intended purpose of the device—amongst other things. In the case of Claude Sonnet, the system prompt suggests it does have an...

Reading: What does it take to ship Rust in safety-critical?, Rust Blog, 14 January 2026. Rust, the programming language, is already used in safety-critical systems. For example, in IEC 62304 one medical device engineer reports: All of the product code that we deploy to end users and customers is currently in Rust. We do...

What makes AI so compelling is not simply faith in technology but the way it suggests we can improve medicine by leapfrogging the difficult work of structural change to confront disease-causing inequality, corporate interests and oligarchic power. Long read from Dr Eric Reinhart in The Guardian (US) from 9 November 2025.

Reading: "Dr. Bot: Why Doctors Can Fail Us—and How AI Could Save Lives", Charlotte Blease.  It's a book looking at the patient-doctor relationship: where it works, where it lets us down, and where AI might fit in.   One section I want to remember describes how Iora Health in the US (now part of One Medical) uses...

I assume it's a difficult problem because I've not found an example of a humanoid robot that can wash itself. This feels odd, given there are impressive sprinting, back-flipping, dancing robots out there. Robots do have different constraints. They could use walk-in baths, UV or chemical cleaning, disposable gloves (or...

This is a hard, perhaps impossible, problem to fix as it stands. Limbic are having a go a it.Reading: The Limbic Layer: Transforming Large Language Models (LLMs) into Clinical Mental Health Experts, PsyArXive Pre-print, 26 August 2025.Limbic have a Class IIa medical device for psychological assessment, and are also...

LLMs do well on medical benchmarks, but this nice experiment blows a hole in that.  Reading: Fidelity of Medical Reasoning in Large Language Models, JAMA Network, 8 August 2025. The set up The experiment presents an LLM with a medical scenario and prompts it to pick the correct answer from a set of 5 options. LLMs do...

Coursera’s AI for Medicine specialism includes a nice, short, set of videos that explain standard model evaluation metrics. I found it useful, and made a few notes. Interpreting accuracy in terms of probability If accuracy of a model is P(correct), we can break that down as the sum of joint probabilities. P(correct ∩...

I’ve been reading an annoying book which uses odds ratio (and hazard ratios). As I don’t use these terms much, I have to keep checking my understanding. Odds ratios: it’s the ratio of two odds! The useful and beautiful The Art of Statistics gives a clear example: Non-bacon eaters have a 6 out of 100 chance of bowel...

Reading: Dedicated AI Expert System vs Generative AI With Large Language Model for Clinical Diagnoses, JAMA Network Open, 29 May 2025. I’d assume an expert system for clinical diagnosis, developed over 40 years, would perform better than an off-the-shelf consumer LLM, like ChatGPT. It does! But only just, and the...