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.
A single-centre pilot put GPT-5 against radiologists on PET staging of oesophageal cancer. Its overall accuracy looks respectable. Its sensitivity for the metastases that change the treatment plan does not — and that is the number that decides whether a tool is usable.
A family used a chatbot to make sense of an MRI report in the weeks before a surgeon could see them. The account is honest and human — and it is one self-reported case, with no comparator and no follow-up. Worth reading for the gap it names, not the method it proves.
A JMIR correspondent piece surveys digital narrative medicine and reports a striking figure: over 90 percent of one antistigma project's participants had head injuries that no chart recorded. It is opinion, not data — and still worth reading carefully.
A recovery tool that about a third of patients found painful to reread. A JMIR feature on digital intensive-care diaries is genuinely moving — and it is journalism borrowing other studies' numbers, not new evidence.
A stepped-wedge trial gave 67 families a web tool to set and track care goals for medically complex children. Nearly all used it in the room; most never touched it after. The signal that survived came from the conversation, not the software.
A viewpoint in JMIR Dermatology argues that for millions, the first dermatological opinion arrives by phone — from someone with no clinical training. It has no new data of its own, but it reads the field honestly.
A benchmark fed three language models deliberately corrupted medical text. Misspellings and sound-alikes barely dented them; blacking words out did the damage. Useful — but it is synthetic noise on a test set, not evidence from a ward.
Twenty junior doctors describe how clinical alerts stop being read. The useful finding is not that they ignore warnings — it is that fatigue is a moving equilibrium shaped by culture and design, not a fixed trait you can configure away.
A small eye-tracking pilot makes an uncomfortable point: the people asked to keep a precise sleep diary are the ones whose attention the poor sleep has already eroded. The interface is not neutral — but this is a pilot, and it measured strain, not cure.
An explainable AI model predicted how long myeloma patients would stay on treatment, using twenty years of Japanese claims data and 647 variables. The discrimination is modest and fairly reported. The part that needs a careful eye is who paid, and which finding they got.
Four NEJM AI podcast interviews, recorded months apart, keep landing in the same three places: a values vacuum, a bias we taught the machine, and a trust gap that tracks consequence. None of it is evidence. The agreement is still worth an hour.
A scoping review screened 275 records to find every AI model meant to prevent surgical complications and follow it to the bedside. Of 19 studies, the models were often accurate. Two are in routine use — and the bottleneck is not the algorithm.
A meta-analysis of machine learning for detecting depression in text screened 3,067 papers and kept 11. The pooled signal is strong — but the prediction interval, from near-zero to near-perfect, is the finding that should travel.
A Chinese single-centre study built a model that flags older surgical patients likely to sleep badly the night before an operation. The discrimination is high — but it was validated on itself, and what it predicts is a label, not a recovery.
A PRISMA review of 65 studies finds language models consistently beat classical methods at classifying clinical text. The honest reading is narrower: it is a synthesis of single-site accuracy studies that mostly never asked whether the models work at the bedside.
A small Stanford study had clinicians and parents rate the same AI-written clinical summaries. They disagreed, significantly — and that disagreement, not the scores, is the finding worth keeping.
A clinical model clears validation, goes live, and slowly drifts — and no one is assigned to watch. A narrative review maps why oversight, not algorithms, is now the binding constraint on healthcare AI. Read for what a review can and cannot prove.
A genuine randomized trial put a medical language model beside nine cardiologists on 107 complex cases. The result is real — and narrower than the headline. It measured preference, not outcomes, and the system was its makers' own.
A study of 326 physicians found pass rates on a clinical-reasoning test rose from 6.4 to 58.6 percent after a 90-minute AI course. The number is real. The design cannot tell us how much of it was the teaching and how much was simply being handed GPT-4.
A single-centre survey of 330 oncology patients found comfort with AI falls steadily as the decision moves closer to life and death — and that the loudest worry is not the machine but the loss of the doctor. A careful read of who was asked.
An experiment had 149 students fact-check health myths with ChatGPT. Misconceptions about flu vaccination fell measurably; those about climate change did not move at all. The asymmetry tells you more than either result alone — and the evidence is thin.
In 118 of 150 blind comparisons, physicians preferred GPT-4.0's answer to a doctor's. The figure is real. But the doctors' answers came from a Reddit forum, and no one ever checked which answer was correct.
A retrospective study grounded a triage language model in two sources at once — a local guideline and three thousand past cases. The honest finding is quieter: it beat an ungrounded model, the authors call it preclinical, and no patient was followed forward.
A Dutch model flags who will be admitted from the emergency department a median 111 minutes before the clinician commits. The number holds. What it measures — decision time, on records, never a patient — is narrower than it sounds.
An analytical synthesis argues that poisoning a medical AI scales with the absolute number of tampered records, not their share of the dataset. The reasoning is sound and worth knowing. But it is a threat model, not a measured event — and that distinction is the point.
Most hospitals evaluate a clinical AI after they switch it on. This qualitative study did the rarer thing: it sat down with the 14 people who would use the system and asked what they expected — and the worries were as telling as the hopes.
An academic centre embedded language models in its medical record and counted what happened: a thousand voluntary users, claimed millions in savings, and — to its credit — two unsupported statements per summary. A candid deployment report, not a controlled study.
When clinical AI disappoints, the reflex is to try a better model. A World Economic Forum essay argues the bottleneck sits one layer down, in the data. The intuition is right. The blueprint is an opinion, and its one hard number proves something narrower.
A meta-analysis pooled studies of AI for fatty-liver diagnosis and reached near-biopsy accuracy. The honest reading is in one comparison: retrospective AUC 0.98, prospective 0.90. The closer you get to real patients, the smaller the edge.
When patients decide whether to use an AI-enabled device, the most persuasive fact is not the accuracy figure or the privacy policy. A survey experiment measured what does move them — and the answer is humbling for anyone who builds the technology.
Anthropic wired Claude to coding systems, trial registries and consumer health apps. It is a real and useful step. But an announcement is not evidence, and a data connection is not a clinical outcome — read it as the application layer it is.
A Taipei team let patients flag their own wound infections through a chatbot. Those who had watched a chronic wound for months agreed with the surgeon almost every time; those days out of surgery did barely better than chance. Experience, it turns out, is a variable.
A urology group structured 1,800 free-text MRI reports through a GPT-4 pipeline at under a cent each. The throughput is real and the plumbing is the point — but the word 'validation' in the title describes earlier papers, not this one.
Fifty studies and 24,764 health professionals, pooled into one acceptance model. Usefulness and technical quality lead. But the link from intention to actual use is the weakest path in the model — and the whole thing measures association, not cause.
A hepatology decision-support system was built to stop answering when its evidence runs thin, and to flag the answers it gives anyway. The architecture is the interesting part. The evidence behind it is thirty questions, scored by its own makers.
A systematic review of 39 medical AI benchmarks finds the same pattern everywhere: models that score 84-90 percent on licensing-style exams fall to 45-69 percent on tasks that resemble clinical work — and to 40-50 percent on safety. The gap is structural, not a fluke.
A purpose-built clinical language model scored AUROC 0.59 at predicting heart-failure readmission. A far simpler embedding, trained on the dataset's own codes, scored 0.65. The more interesting number is that neither is good enough to act on.
Three in four people who told a chatbot they felt low were not asking for advice — they were asking to be heard. A formative study of eight commercial systems shows most answered with information instead, and names the gap precisely.
A US health system taught software to read scanned colonoscopy reports and write follow-up dates into the record unsupervised. The build is clever and honest. But it is a single-site proof of concept, and only about a third of reports ever reached the automated step.
In a real colorectal-cancer dataset, the single most prognostic variable — tumour stage — was absent from 75 percent of records, and half the rest were miscoded. A quiet, careful paper on why a model can only learn what the data actually contain.
A survey of 218 primary-care patients found that those with visual disabilities used voice assistants less often than everyone else — yet relied on them far more heavily when they did. A small, careful study with a finding worth sitting with.
A geospatial study of 3,092 US hospitals asked not whether predictive AI works but where it goes. It pools in the better-connected, better-resourced places — and the top predictor was interoperability, not size or money.
A doctored referral letter, an embedded instruction in a PDF, a Unicode trick in the OCR output—the next generation of cyberattacks on hospitals no longer targets your firewall, but your AI.
The largest survey yet of hospital patients' views on medical AI — 13,806 people, 43 countries — finds the headline is not enthusiasm. It is two conditions: show me how it reasons, and keep a physician deciding.
A JMIR viewpoint reframes the FHIR-versus-OMOP-versus-openEHR debate. The technical specification matters less than the open-source community around it. The argument is sound; it is opinion, not evidence, and the authors have skin in the game.
A feasibility pilot put volunteer medical translators one video call away from the bedside. Over two months it logged 39 requests and connected on 16 of them. The honest finding is in the 23 that went unanswered — and in what the study never measured.
Wiring a language model to a curated medical knowledge graph is meant to stop it inventing diagnoses. A peer-reviewed study tested it, and the honest result is a trade: better reasoning, worse abstraction.
A peer-reviewed scoping review maps the gap between signing up to responsible-AI principles and being able to enforce them. Its contribution is a vocabulary for that gap — not an implementation that has been shown to work.
For a century we paid for the answer and tolerated the manner. As machines make the answer cheap, the scarce thing left is the part we always undervalued: getting a frightened human being to actually follow the plan.
We treat a clinical note as thorough only if a doctor typed every word of it. We have confused the pain of writing with the quality of the record — and we keep rewarding the wrong thing.
A resident scribbling potassium values on the back of their hand is not careless. They are out of working memory, and a ballpoint pen on skin is still a better interface than the record we built them. A case for automating the bureaucracy so the mind is free to reason.
An AI that reads a scan and an AI that summarises a history are not the same kind of machine, and they do not earn our trust the same way. One we validate against trials. The other we must be able to check, line by line, against its source.
We hire some of the most expensive minds in the economy and spend half their day on tasks that require literacy, not a medical degree. That is not a wellness problem. It is a misallocation of capital we would never tolerate anywhere else.
Show a hospital a problem and its first instinct is to add something to the room: a tablet, a sensor, a robot in the lobby. The harder, better question is what we could take away. A case against electrified clutter.
For a century, medicine rested on a quiet asymmetry: the doctor held the facts, the patient did not. That asymmetry has dissolved. A patient with a phone now reaches the same survival curves as the professor. So what is the doctor for?
We track bed occupancy, cash flow and medication stock to the decimal. The one resource we never measure is the one most likely to run out on a ward: the attention of the person at the bedside. What we decline to count, we quietly allow everyone to spend.
Hospital dashboards worship speed: door-to-doctor, length of stay, throughput. We buy technology to remove every pause between symptom and prescription. But in medicine the pause is not a defect in the line. It is the work.
Anxious young doctors spend their weekends learning to code, sure it is how they survive the AI wave. A rational reaction to a failing curriculum, and the wrong problem. The skill that matters is not building the tool but knowing when it lies to you.
Every patient in a modern hospital has a double: a digital profile assembled to fit the reimbursement rules. Our safety systems increasingly consult that double rather than the person in the bed. When the codes are wrong, the consequence is no longer just an audit risk.
We built clinical software around a patient who does not exist: one clear complaint, one tidy pathway. Real patients arrive in the plural. When the form will not bend, the clinician bends the truth instead — and clicks the least wrong box.
When old software failed, it failed loudly — garbled text you could not miss. Generative AI fails in flawless prose. The error no longer announces itself, and a perfectly written note may be the most dangerous thing in the record.
The patient moves from the GP to the specialist to the ward and back. Their record does not. Every threshold they cross, they leave a part of their history behind — and we treat the resulting blindness as a plumbing problem rather than a safety one.
For a century we selected doctors for what they could memorise, because the physician was the storage medium. That problem is solved. The skill that remains scarce is the one we barely test for: making sense of a patient when the data disagrees with itself.
Healthcare technology is sold on minutes saved. But a minute spent logging in at eight in the morning is not the same as a minute spent deciding on an end-of-life plan at four in the afternoon. The scarce resource is not time. It is judgement.
Watch a doctor on rounds and the frustration is rarely about the medicine. It is about the screen, which interrupts clinical thought to confirm a billing code. We did not digitise the chart. We digitised the invoice and taped the patient's history to the back.