For how we live.
A January 2025 draft guidance lays out the documentation the agency expects across an AI device's whole life, including bias checks across demographic groups and postmarket monitoring. It is not yet binding.
A deep-learning model reads the pancreas on non-contrast CT taken for other reasons. It reached very high accuracy in large tests — none of them a prospective screening trial.
Software that interprets cardiotocography in labour lifts clinicians' accuracy in reader studies. The largest randomised trial of computerised interpretation, in 47,062 women, found no benefit for mothers or babies.
The Hypotension Prediction Index warns anaesthetists before blood pressure falls. Three 2026 trials find a plain mean-pressure alarm does the same job, and the algorithm's benefit tracks how hard clinicians treat.
A final order files AI tools that flag possible cardiovascular disease as class II with special controls, letting similar triage software clear via 510(k). The controls require real-world, subgroup-tested validation.
A predetermined change control plan is the FDA's answer to a model that keeps learning: pre-approve a bounded set of future changes, and the manufacturer can ship them without returning for a fresh review.
Systems that give a diagnosis with no clinician in the loop post high sensitivity in trials. In real-world use, one in six images was too poor to grade — and the strongest numbers come from company-run studies.
Pooled diagnostic accuracy is high in specialist settings, but performance falls on Fitzpatrick IV–VI skin and on smartphone images, and most studies never report skin tone at all.
A deep-learning system drew tumour targets and organs at risk automatically. It saved time and improved consistency — but oncologists still reviewed and edited the contours before treatment.
A Polish study found endoscopists' unaided polyp-detection fell after AI arrived. A Norwegian trial that tracked doctors before, during and after AI use found no such deskilling.
Algorithms that read short voice recordings can pick up acoustic signs of some conditions. In the largest and most careful studies the signals are real but narrow, sex-specific, and often fail to generalise.
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.
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.
A hamstring analysis separated pre-injury from return-to-sport sprint mechanics at AUCs near 0.79. A multi-team rugby model tried to predict injuries before they happened, and could not.
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 secondary analysis of 8,376 adolescents built two models to predict who benefits from school-based mindfulness training. Both found a subgroup — and the difference in outcomes was trivial.
Trained on 2,202 retracted paper mill articles, a text classifier flagged 9.87 percent of the cancer literature — including in the highest-impact journals. Flagged is not the same as fraudulent.
A January study forecasts particulate pollution across six Gulf states and reports Qatar highest at 87.90 µg/m³. The modelling is careful; the input data is annual, national and thirty-one points long.
A transformer model trained on 17,119 scans beat expert and non-expert examiners on every metric and cut simulated referrals by 63%. The catch is who built it and the validation the rest of the field still lacks.