The Health AI Brief

Why do some AI models get 99% accuracy in a study but fail in the real world? The answer is "overfitting" – a critical flaw where a model memorises data instead of learning from it. In this episode, we use the analogy of an over-crammed student to explain what overfitting is and give you the key question you must ask to spot it in any research paper. #Overfitting #AIinHealthcare #MachineLe...

Paper - Human-level control through deep reinforcement learning, by Mnih et al 2015, Nature Link - https://www.nature.com/articles/nature14236 This week on The Health AI Brief, we're going back to basics. We revisit the landmark 2015 DeepMind paper that taught an AI to master 49 different Atari games from scratch. Find out why this wasn't just about video games, but a foundational breakthr...

What happens when you combine an AI that can 'do' with an AI that can 'judge'? You get an Actor-Critic partnership, the state-of-the-art in reinforcement learning. We explain how this 'trainee and consultant' model is powering the next generation of dynamic, responsive medical AI. #HealthAI #DigitalHealth #ArtificialIntelligence #ReinforcementLearning #DeepLearning #Clinica...

Forget judging, some AIs learn by doing. 'Policy Gradient Methods' create an 'AI Actor' that learns skills directly, much like a person learning to suture. This is the technology behind AI in robotics and precision medicine. Here's what you need to know. #HealthAI #DigitalHealth #ArtificialIntelligence #ReinforcementLearning #RoboticSurgery #ClinicalAI #MedEd #MedicalPodcast #a...

What happens when a patient decides an AI chatbot is more "humane" than their doctor? A powerful essay in The Guardian and Rest of World explores one mother's growing reliance on 'Dr. DeepSeek', an AI that is both a comforting, empathetic companion and a source of dangerous medical misinformation. In this episode, we dissect the story's key themes: the "push" from a...

Some AIs learn by becoming expert judges, calculating a score for every possible clinical decision before making a move. We explain value-based methods, the 'AI Critic,' and why they excel at multiple-choice medicine but falter when the decisions are infinitely complex. #HealthAI #DigitalHealth #ArtificialIntelligence #ReinforcementLearning #DeepLearning #ClinicalAI #MedEd #MedicalPodcast...

Epic recently unveiled Comet, a new AI model trained on 118 million patient records to predict future health events. The scale is unprecedented, and its initial ability to outperform specialised models is a huge leap forward for clinical AI. But what is it really learning from our messy, real-world data? In this today's episode, we break down why Comet is a landmark achievement but also an imp...

You can't teach an AI complex medicine by throwing it in at the deep end. Curriculum learning applies the principles of medical school to AI, training models on simple tasks before moving to complex ones. Find out why this matters for building safe and effective clinical AI. #HealthAI #DigitalHealth #ArtificialIntelligence #MachineLearning #CurriculumLearning #ClinicalAI #MedEd #MedicalPodcast...

An AI, like a clinician, faces a constant choice: stick with the proven treatment or explore a novel approach? In this episode, we break down the 'exploration-exploitation' dilemma, a core concept in AI that has major implications for how we design and trust medical AI systems. #HealthAI #DigitalHealth #ArtificialIntelligence #ReinforcementLearning #ClinicalAI #MedEd #MedicalPodcast#ai in...

Paper: Scaling Large Language Models for Next-Generation Single-Cell Analysis by Rizvi et al Paper link: https://www.biorxiv.org/content/10.1101/2025.04.14.648850v2 This week we're covering recent research presenting C2S-Scale, a new model from researchers at Yale and Google that teaches Large Language Models the "language of the cell." By translating complex genomic data into "cel...

How does an AI like ChatGPT learn to be so helpful? The answer is "Reinforcement Learning," a powerful method of learning through trial-and-error, rewards, and punishments. In this special extended episode, we break down how reinforcement learning works and explain RLHF, the key technique used to train the language models that are transforming our world. #ReinforcementLearning #RLHF #AIinH...

We are rolling out powerful AI tools in hospitals and clinics at a breathtaking pace. But are they helping, or are they causing harm? A new report from the JAMA Summit on Artificial Intelligence reveals a key gap in our ability to answer that question. Featuring a stark warning from former FDA Commissioner Robert Califf, this episode breaks down why we are "flying blind" and explores the r...

Why build an AI model from scratch when you can give it a head start? And why waste expert time on easy cases? This episode explores two powerful strategies for efficient AI development. Discover how Transfer Learning gives your model a foundation of pre-existing knowledge and how Active Learning creates a smart feedback loop where the AI asks for help on only the toughest cases. #HealthAI #Medica...

How can an AI learn to read a medical scan without a perfect, expert-labeled dataset? In the real world, data is messy. This episode dives into three ingenious techniques (semi-supervised, self-supervised, and weak supervision) that allow AI to learn from a little bit of expert guidance, teach itself from unlabeled data, or make sense of noisy, imprecise information. #HealthAI #MedicalAI #SemiSupe...

The New England Journal of Medicine just featured an AI, 'Dr. CaBot,' as a guest expert in its legendary diagnostic challenge. This AI can not only find the right diagnosis but can reason and tell a compelling clinical story, sometimes more convincingly than human doctors. But does this mean Dr. AI is ready for the ward? We explore the gap between a perfect, curated case and the messy real...

What if an AI could find patterns in patient data that we've never seen before? That's the power of "unsupervised learning", a type of AI that learns without an answer key. In this episode, we explain how this method works, and why it's a powerful tool for discovering new patient subtypes and advancing personalised medicine. #UnsupervisedLearning #AIinHealthcare #MachineLearnin...

How do we teach an AI to read an ECG? The most common method is "supervised learning," which is a lot like using flashcards with a medical student. In this episode, we explain this fundamental concept and reveal the two critical questions you should always ask about the data to assess the quality of any medical AI model. #SupervisedLearning #AIinHealthcare #MachineLearning #ClinicalAI #Hea...

Top AI models are acing medical benchmarks, but are they actually ready for the clinic? A groundbreaking study reveals that impressive scores can hide a dangerous lack of real-world robustness. In this episode, we break down the ingenious "stress tests" that expose how AI can succeed on an exam for all the wrong reasons—from guessing answers without seeing medical images to failing when th...

An AI model doesn't just learn on its own; it follows a protocol. The settings of that protocol, like the "learning rate", are called hyperparameters. In this episode, we explain what these crucial settings are, why they are the 'art' of AI development, and how they help you judge the quality of a research paper. #Hyperparameters #AIinHealthcare #MachineLearning #ClinicalAI #He...

Imagine trying to find the lowest point in a valley while blindfolded. How would you do it? The same way an AI finds the best answer: one step at a time, always moving downhill. This process is called "gradient descent," and it's one of the engines that powers machine learning. In this episode, we explain how it works, what the "learning rate" is, and why it matters for underst...

The UK has just launched a star-studded National Commission to rewrite the rulebook for AI in the NHS. The goal: faster, safer innovation for patients. It could be a powerful accelerator and will hopefully avoid the pull of becoming another talking shop lost in bureaucracy. #HealthAI #AIinHealthcare #DigitalHealth #NHS #HealthTech #Regulation #MHRA #Innovation #MedTech #PatientSafety #FutureofHeal...

How does an AI model quantify a mistake? It uses a "loss function" – a scorecard that penalises different types of errors. In this episode, we explain what a loss function is, why it's not a one-size-fits-all tool, and how it reveals the true clinical priorities of any AI model. A crucial concept for critically appraising new research. #LossFunction #AIinHealthcare #MachineLearning #Cl...

How does an AI model actually learn to spot disease on a scan? It all comes down to one fundamental goal: minimising error. In this episode, we kick off a new set of episodes on the mechanics of machine learning by explaining this core principle with a simple clinical analogy that will change how you look at AI. Understanding this is the first step to critically appraising any research paper that...

Shmatko, A., Jung, A.W., Gaurav, K. et al. Learning the natural history of human disease with generative transformers. Nature (2025). Link to paper: https://www.nature.com/articles/s41586-025-09529-3 What if an AI could forecast your health like the weather? A groundbreaking new model called Delphi-2M, published in Nature, claims to do just that — predicting your risk for over 1,000 diseases using...

We break down a landmark UCL study on the NHS's £21m programme to deploy AI in chest diagnostics. They uncover the real reasons for significant delays, moving beyond the technology to the critical, real-world barriers: staff capacity, fragmented IT infrastructure, and complex governance. Find out why dedicated project management is the secret to success and what this means for the future of AI...