"UCL Hospitals uses AI to predict patient deterioration"

Weekly Updates: Clinical AI Academy (13–17 October 2025) Here is what happened last week [6–10 October 2025]: Guidelines and standard frameworks for artificial intelligence in medicine: a systematic review. https://lnkd.in/eq_qH22t A Current Review of Generative AI in Medicine: Core Concepts, Applications, and Current Limitations. https://lnkd.in/eTngZTnt Clinical AI Rounds: What is AI-normal? https://lnkd.in/ehPBJQZC Here is the plan for this week [13–17 October 2025]: Discussions and papers: The Joint Commission and Coalition For Health AI (CHAI) Guidance on Responsible Use of AI in Healthcare. The Illusion of Readiness: Stress Testing Large Frontier Models on Multimodal Medical Benchmarks. More Clinical AI Rounds. Join us: https://lnkd.in/evemjjXu

🗞️ Out now: CODEX Digest - October 16, 2025 This week's digest features a comparative study exploring the use of AI to identify diagnostic uncertainty in real time, new reporting guidelines for diagnostic accuracy using AI, and an analysis of clinicians' perceptions of generative AI use in medical decision-making. Also highlighted this week is a look at how medical educators provide negative feedback when the clinical outcomes are adverse and recommendations for medical learners using AI in the emergency department. 👉 Check out this week’s digest: https://lnkd.in/gSZM5ZZY 📬 If you haven't already, subscribe now to get it straight to your inbox every Thursday: https://lnkd.in/gAZhHQek

Hospitals need assurance before deploying AI, but validating every model extensively isn’t sustainable. The result is slow adoption, duplicated effort, and uncertainty about what’s truly safe to use. AI validation doesn’t have to be an all-or-nothing effort. Between blind trust and endless testing lies a smarter path that adjusts validation depth to actual risk and the evidence behind each algorithm. Join Dr. Franz MJ Pfister, MD, MBA. Franz MJ Pfister (deepc) and Kicky van Leeuwen (Romion Health) present a framework to help radiology teams make validation proportional, practical, and trustworthy. Beyond Accuracy: A Framework for Trustworthy Radiology AI Validation 16 October 2025, 6:00 pm CET 👉 Register to join the live session https://lnkd.in/d3iifyZe

See how generative AI and clinical terminology are transforming systematic literature reviews to make evidence generation faster and scalable. https://bit.ly/4heJg4l IMO Health

Interesting research on how guessing is embedded in large language models and on the fragility of the benchmarks methods used so far. "Leading systems often guess correctly even when key inputs like images are removed, flip answers under trivial prompt changes, and fabricate convincing yet flawed reasoning. These aren't glitches; they expose how today's benchmarks reward test-taking tricks over medical understanding." #AI #LLMs https://lnkd.in/gZFrUbBS

A growing number of papers confirm essential 𝐠𝐚𝐩𝐬 𝐛𝐞𝐭𝐰𝐞𝐞𝐧 𝐀𝐈 𝐚𝐧𝐝 𝐡𝐨𝐥𝐢𝐬𝐭𝐢𝐜 𝐡𝐮𝐦𝐚𝐧 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞, but this is probably just the surface of the whole truth. Challenging AI by focusing on its errors and comparing it with holistic human intelligence can help move AI forward. I believe that turning attention to the core building blocks and challenging underlying beliefs—especially the analogy with the human brain and neurons -> can shed light into the darkness. Simply adding more data, more reasoning examples, and more GPUs is like scaling up damaged neurons. Current AI models feels for me like an smart student who red whole internet but never made an real experience. Where do you think the essential focus of AI research should be? https://lnkd.in/dSyKshZ8 https://lnkd.in/dp9SjZfs

I was a guest speaker at an interesting session organized by Oslo legeforening (Oslo doctors' association?) a while back. One of the other talks was about how they had developed an in-house AI-model for recognizing brain hemorrhaging and then compared it to commercial models. Their own model performed at about the same level. But the most interesting fact to me was that cost-wise, using commercial AI tools  to scan medical images was MORE EXPENSIVE than using human radiologists.  So could we save money by using generic LLMs instead? While they're being funded by the hype machine, their costs are bearable at least.  This image is from a study about how generic AI models compare to human radiologists. (Link to the study in the comments) Turns out the generic LLMs cannot be used safely. The best model only succeed 30% of the time.  Lesson seems to me that if you want to use AI in medical scans: build and maintain your own models.   

Does Multimodal AI outperform Unimodal AI in medicine? A recent scoping review titled “Navigating the Landscape of Multimodal AI in Medicine” analyzed 432 papers (2018–2024) and found that multimodal AI models consistently outperform unimodal ones, with an average gain of 6.2 percentage points in AUC. However, this progress brings new challenges, from data alignment and fusion design to handling missing modalities and the limited availability of large, diverse public datasets. The paper provides a clear overview of multimodal AI across medical fields and data types, highlighting the key factors driving progress and outlining the steps needed to move the field forward. 📄 Read the full paper here: https://lnkd.in/dFGp_FAX #MultimodalAI #DeepLearning #AIinHealthcare #MedicalAI #DataIntegration

If you cannot explain why an AI made a critical clinical recommendation, you should not deploy it. The standard in medicine is transparency. The #AIinHealthcare black box must be cracked: we need Explainable AI (XAI) to foster trust and enable physician oversight. #ExplainableAI #ClinicalSafety _____ https://lnkd.in/gtaKrEku

I was lucky enough to attend a talk by Dr Jeremy Wyatt on the benefits of AI in healthcare at a local Univeristy of the Third Age conference. It was very insightful and some interesting points were made. AI’s ability to enhance diagnosis through analyzing medical data and finding patterns was particularly interesting. Following this talk I decided to research the use of AI in rural medicine settings for a possible EPQ topic. (Image from another speech at the same conference)