For how we live.
AI, devices, digital health, and what the evidence supports.
Linking exome data to electronic medical records solved 15% of unexplained hearing-loss cases and flagged new candidate genes — while showing what the records could not supply.
In 216 patients with chronic patellofemoral pain across 10 Korean hospitals, an 8-week digital therapeutic beat education and exercise instruction on pain — by 12.5 mm on a 100 mm scale.
A Danish stepped-wedge trial gave endoscopists automated feedback on their technique after every procedure. Adenoma detection rose from 43.4% to 48.6% — a different tool from real-time polyp AI.
Trained on 1.7 million ECGs paired with clinicians' report text, ECG-CLIP reached the same accuracy as the best comparator with about 90% less training data — a bid at the field's labelling bottleneck.
In 6,772 users of a US health system's AI triage tool, people engaged about twice as often when the AI matched what they already planned — raising a hard question about what the tools are steering.
A review of 41 studies found that the great majority of AI medication-adherence prediction models carried high risk of bias, and that fancier algorithms did not reliably predict better.
Relationality, self-governance, competence. The position paper's starting premise is that consensus on privacy, disclosure and fairness has not been reached, and clinicians need guidance anyway.
Across 24 official German licensing exams, the best model answered 99.31% of first-exam items correctly. On items containing an image, the error rate rose several-fold, against 1.24x for students.
Thirty-two interviews across Uruguay, Argentina, Chile and Mexico find implementation turned on governance, not software. A separate patent analysis shows where digital therapeutics are not being built.
A network meta-analysis of 53 studies and more than 7 million admissions finds machine learning out-discriminates traditional sepsis scores — and would raise roughly two false alarms for every real one.
A system called Quicker writes clinical guideline recommendations through a GRADE workflow. A Matters Arising and its reply, published the same day, map what still has to be built around it.
European hypertension specialists found only two wrist-cuff blood pressure watches with validation studies done to established protocols. A sleep group ran a Galaxy Watch against two nights of polysomnography.
A researcher who expected the evidence base to be thin says even she was surprised by how thin. Most cleared devices never appear in a registered clinical trial at all.
The tool reads vital signs, lab results, and history the way clinicians already do — just faster and continuously. Its accuracy tops out around 90%, which means it's also still wrong a meaningful share of the time.
A new discussion paper proposes a two-axis risk framework and a physician-training analogy for evaluating generative AI devices. It is not a rule, and the agency is asking what one should look like.
Emulating a trial inside Veterans Health Administration records, researchers estimated adjusted mortality risk ratios of 0.90 to 0.84 for starting continuous glucose monitoring in older-onset type 1 diabetes.
It found no meaningful difference at six weeks. It enrolled 64 people, tested one device, and ran for six weeks. That is the entire head-to-head evidence base three years after the category opened.
A Nature Reviews Drug Discovery audit says the problem isn't the models. They were built to be validated rather than used, and benchmarked against the wrong thing.
START-AI reads triage comments, case notes and whether anyone ordered a blood test. Adding heart rate and blood pressure to the model produced no measurable improvement — a result the team published rather than buried.
A study of 2.4 million veterans a month finds health information exchange volume moving outcomes in opposite directions depending on which side of the exchange you are on.
A blinded comparison of 180 responses from a university telehealth centre in Minas Gerais found AI matched human specialists on medical adequacy and risk, and beat them on comprehensibility.
42 orthopedic physicians diagnosed 40 rare diseases twice, once alone and once after seeing AI suggestions. Accuracy jumped 20 to 26 points, but the same-day design leaves memory unaccounted for.
A Lancet Digital Health scoping review found the field's fairness metrics fragmented and rarely clinically validated. A second review found that most studies don't measure fairness at all.
A review of Paige Prostate Detect and Ibex Prostate Detect finds the tools mainly help less-specialized pathologists, and warns performance shifts when a tool meets a new population.
Maccabi Healthcare Services studied 626 of its own physicians to work out who follows algorithmic prescribing advice — and found the pattern was about practice structure, not just attitude.
A paper in the journal Resuscitation lays out how a large language model could help dispatchers recognize cardiac arrest and coach CPR in real time — while acknowledging the concept still needs to prove it saves lives.
A small proof-of-concept study pitted several AI systems against gastroenterologists and emergency physicians on cholangitis exam questions. The gap between the best and worst AI performers was enormous.
A new structured dataset finally makes the approvals countable. Two trials published weeks apart show why counting them says little: 0.86 fewer migraine days in one, a 20-point symptom shift in the other.
The agency granted its first authorization for a software-aided adjunctive diagnostic device in wound assessment, a category built for tools that analyze a wound optically.
A study running 5.3 million evaluations through nine large language models found eligibility judgments were largely stable across identity labels — except when a patient vignette mentioned homelessness.
A cross-sectional analysis of NICE evaluations found 78 supporting studies behind 30 technologies — and consistent holes in comparators, cost of delivery and adverse-event reporting.
DxDirector-7B beat human physicians on a benchmark of complex diagnostic cases while requesting far fewer tests. Its own authors say it isn't ready for high-risk or emergency cases.
A secondary analysis of the VITAL-AF trial reports a screening benefit only in the top risk decile — with a confidence interval whose lower bound sits at 0.01.
A Nature Communications review of the continent's biomedical data science capacity found none of 36 surveyed groups using cloud high-performance computing, and named electricity outages among the most common obstacles.
GE's Critical Care Suite gains an algorithm that flags misplaced enteric tubes on a chest X-ray — a complication that reviews of the practice have linked to respiratory harm and, in some cases, death.
A content analysis of every FDA cybersecurity safety communication from 2013 to 2025 finds a small corpus, rising over time, and 94% of the flagged vulnerabilities rated severe.
The clearance is the third in a growing family of machine-learning notification algorithms built to spot serious cardiac conditions from a routine 12-lead ECG, a test most patients already get.
A new audit of China's regulatory record finds a market concentrated in radiology, dominated by deep learning, and clustered in four cities. It also finds the approval curve flattening.
A 1,298-person randomized study found people using chatbots to work through medical scenarios did no better than people without them — even though the same chatbots, tested alone, got the right answer most of the time.
Researchers ran 960 responses through ChatGPT Health using clinician-written vignettes. Failures clustered at both extremes, and crisis safeguards activated unpredictably.
Three February papers push plasma p-tau217 past detection: clock models estimating years to symptom onset, four assays compared head to head, and a real-world audit of what the result changes.
A February theme issue documents faster notes, happier clinicians and enterprise rollouts to thousands. It also contains an editorial asking the question none of the studies answer.
A 250-person randomised trial paired a wearable tracker with just-in-time energy-management messages. Post-exertional malaise fell in both arms, and the difference between them was trivial.
Ten studies, wide confidence intervals, and factual error rates of 26 to 36 percent in AI-drafted documentation. The review's own conclusion is that the evidence is preliminary and highly uncertain.
A trial across five Chinese hospitals found remote robotic urological surgery non-inferior to local surgery. The sample is small, the surgeons were experts, and the network never failed.
Clinicians rated empathy at 4.6 out of 5 and quality of information at 2.7. Every chatbot produced at least one piece of guidance judged inappropriate, overstated or inaccurate.
An 11,018-measurement study in 24 English intensive care units found false negative rates up to 35.3 percentage points higher in patients with darker skin tones. A neonatal study published two days later did not.
The announced lab pairs robotic wet labs with computational models in a continuous loop. It is a serious bet on a real bottleneck — and, so far, entirely a forward-looking statement.
Revised guidance reverses the 2022 position that a single recommendation made software a device. The agency's own criterion — that a clinician can independently review the basis — now carries the weight.
The FDA's revised general wellness policy lets non-invasive devices infer blood pressure, glucose and other physiologic parameters. The regulated line is no longer the measurement — it is the sentence next to it.
The EAGLE trial ran colonoscopy AI off-site over a network and aimed it at the lesions that matter. It reports a threefold gain in serrated lesions — in a field whose main US guideline recommends nothing.
One agency asked how to measure AI devices after they are deployed. Another asked how to speed adoption up. The offices that would answer the first question are losing staff.
Germany's DINKS trial is one of the largest randomised tests of a reimbursable digital therapeutic. The effect is big, the comparator is usual care alone, and the sponsor made the app.
The headline finding is noninferiority. The more useful finding is what the shared denominator was — and how wide a gap the trial was designed to tolerate.
England ran the closest thing to a national test of clinical AI anyone has published. The software sites gained more, but the comparison sites were improving on their own, and that gap is the whole result.
The WISeR model starts on 1 January in six states. The contractors running it are paid out of the savings their determinations produce — the design feature doctors keep pointing at.
Two November trials found ambient AI cut documentation time and work exhaustion. A parallel emergency department comparison found physicians spent more time in the note, not less.
A systematic review of every AI device authorised from 1995 to June 2024 found 97% cleared through the 510(k) pathway, which does not require independent clinical data on performance or safety.
Ten sonographers estimating gestational age got sharply more accurate when shown model predictions. Adding visual explanations improved the average further but made several individuals worse.
Physicians editing AI drafts wrote better notes than they did unaided, and worse notes than the AI produced alone. Both directions of that result matter.
PpgAge predicts chronological age from consumer wearable photoplethysmography and tracks disease. The same optical method has documented accuracy problems across skin tones.
Twenty experts scored three language models on ten common questions. One model cleared the content validity bar on both topics; another scored zero on jet lag.
Across 40 trials, self-guided apps moved panic severity modestly and barely touched fear-related cognitions. Clinician-guided versions moved both — which complicates the scalability pitch.
A census of 691 AI-enabled devices cleared through 2023 found demographic data missing almost everywhere, three summaries reporting patient outcomes, and 113 recalls driven mostly by software.
Across 90,670 virtual encounters at one US health system, phone visits scored lower on likelihood to recommend. A separate VA cohort of 1.2 million found telehealth cut same-day mental health access sharply.
An EHR-embedded model wrote hospital summaries residents edited less — but with more confabulations. A separate pipeline audited 21,041 trial reports against CONSORT at 91.7% agreement with experts.
EyeFM was tested as an assistant to 16 ophthalmologists screening 668 high-risk patients in China. Patients in the AI arm also followed referral advice more often.
Google's PH-LLM beat sampled human experts on multiple-choice tests but only matched them on real cases. A new reporting checklist published the same month explains why such claims are hard to compare.
Off-the-shelf features from two published models matched retinal scans to the right patient 78% to 86% of the time. Fine-tuned for the task, one model reached 99.5% on OCT.
Singapore General Hospital randomised residents to work with and without an LLM assistant. Documentation time fell by 1.82 minutes, which was not significant. The economic model used the point estimates anyway.
EchoNext scored 77.3% accuracy on a 150-ECG set where 13 cardiologists averaged 64.0%. Its authors released the model weights and a 100,000-ECG labelled dataset alongside the paper.
A draft-reporting model cut documentation time 15.5% across 23,960 radiographs without changing report quality. A separate reader study found AI assistance raised prostate MRI accuracy by 3.3 percentage points.
Rentosertib was safe over 12 weeks in 71 patients with idiopathic pulmonary fibrosis. A lung-function signal appeared at the highest dose, in a trial not designed to prove efficacy.