Germany's framework agreement on discharge management requires a discharge letter on the day of discharge, and ePA filing has carried sanctions since 2026. Why the deadline routinely fails on data that was never structured at admission.
Over 120,000 patient records exfiltrated at a billing service provider; one hospital is demanding damages. The facts of the unimed incident — and the checklist that follows from them.
An AWS region is built to survive losing one of its three availability zones. In the UAE, two went down at once. The reporting is solid; the lesson the German data debate draws from it is the part worth checking.
A red-teaming study put a patient-facing chatbot under sustained pressure. It never misquoted its documents. It abandoned its own safety rules when a distressed user simply kept asking. Single-question tests would never have seen it.
An open-source team trained codon language models across 25 species for the price of a weekend, and found that starting from an English-text model made things worse. It is a careful engineering write-up — not a clinical result, and not yet peer-reviewed.
Accuracy on a benchmark is not the same as helping a patient, and almost no one has measured the second thing. A careful piece of reporting holds that line — and the two trials it leans on show why it matters.
A heart-rate model built from a US pandemic dataset flagged 42 malaria cases in Kenya that would otherwise have gone unseen. The catch: those 42 alerts rest on data the model invented to fill the gaps.
Run a 3-billion-parameter open model and GPT-4o on two speech datasets and they trade the lead. That reversal, not the headline win, is the most instructive thing in this careful evaluation.
A hospital-association survey finds nine in ten German hospitals have begun connecting to the national patient record, but only a quarter can use it everywhere. The interesting number is the distance between the two.
An AI healing index flags a stalling wound a week before the standard ruler does. The lead time is real and modest — and the study was written by the company that sells the index.
A study of 4,244 people with Long COVID and related conditions asked whether a morning phone reading of heart rate foretells an evening crash. The within-person signal is real. The gain over knowing yesterday was bad is a few points.
The largest real-world cohort yet for a consumer blood-and-wearable platform reports that most users with poor baseline markers improved. The numbers are real. With no comparison group, what they cannot tell us is whether the platform is why.
An international team led from MIT built a prompting layer that makes a language model interrogate its own confidence before it answers. The peer-reviewed version now carries numbers. They are large, and they were measured on synthetic cases scored by a model — not on patients.
A clinical model that gives a wrong answer confidently is dangerous, because people defer to confidence. This study tests a prompt that forces the model to ask first. The behaviour shift is dramatic; the evidence is a benchmark, not a bedside.
A 0.05 fall in a model's discrimination should force someone to re-examine it. In most regimes, no rule names that threshold or whose job it is to act on it. An npj Digital Medicine Perspective proposes one — as a reasoned argument, not evidence.
A venture firm argues that AI has made individuals far more capable while their organisations stayed flat — and that the fix is structural, not another chatbot. The argument is sharp and worth reading. It is also opinion, from people who profit if you believe it.
A clinical AI retrieved every fact it was asked for and two physicians flagged nothing unsafe. Both results are real. Both come from ten de-identified patients on public research data — which is exactly what proof-of-concept means.
At HIMSS 2026 the vendors shipped clinical AI agents faster than anyone could count them. A conference dispatch is not a study — but it shows exactly where the validation gap sits, and how a single word keeps the proof bar low.
An insulin pump scored 8.8 out of 10 for a wireless flaw. A short npj Digital Medicine Comment uses cases like it to argue that device cyber risk is patient-safety risk — and that regulators assess the wrong unit.
A BBC Future feature gathers three of the most persuasive AI-in-medicine stories. Read closely, they share one structure: AI is fast at the front of the pipeline and absent from the slow, expensive back. That distinction is the whole appraisal.
A new perspective proposes that a measured signal, not the calendar, should decide when a multiple sclerosis patient is seen. The triage logic is sound. What is missing is anyone who has tried it.
Twenty clinicians explain why good mental-health apps never reach patients. The obstacle is almost never the technology. It is whose job it is to introduce the tool, watch the alerts, and answer when something looks wrong — questions no software answers.
Forty-nine studies suggest disturbed sleep and activity shadow cognitive decline by years. Only three tested their model outside the lab that built it. The signal is real; the case that it works as a screening tool is not yet made.
Twenty-eight machine-learning models claim to predict delirium after heart surgery. Pooled, they look clinically useful. Read the validation methods and the heterogeneity, and the single number stops meaning what it appears to.