AI in Medicine Watch — Sep 11, 2026

Welcome to your weekly update in AI in Medicine.

This period gives us something the field has badly needed: randomized trials. Two of them, and they point in different directions — a large language model assistant in Kenyan primary care that changed nothing, and a narrow imaging tool in Hong Kong that cut unnecessary referrals by a huge margin. Alongside that, two prediction models that find high-risk patients our current standards miss entirely, one for sudden cardiac death and one for hip fracture. And a set of papers about the humans in the loop — who benefits from an explanation, who gets anchored by one, and what a health system owes its workforce before it deploys any of this. If you are deciding what to buy, build, or trust this year, this period matters to you.

Let's start with the trials, because randomized evidence on AI in live clinical care is still rare enough to be an event. In Nature Medicine, Agweyu and colleagues ran a pragmatic cluster-randomized trial across sixteen primary care facilities in Kenya, randomizing a hundred and three clinical officers to use their electronic medical record with or without large language model assistance, and enrolling nearly ten thousand patients. The primary outcome was an expert-adjudicated composite of treatment failure within fourteen days. The finding is a clean null. Treatment failure occurred in roughly two percent of patients in each arm, with no significant difference, and the direction of effect, if anything favouring the AI arm, was not statistically distinguishable from chance. Crucially, the trial found no safety signal — independent adverse event review turned up nothing attributable to the intervention. So the honest read is that generative assistance in this setting was safe but did not improve outcomes, and any real benefit is probably modest. Strengths here are the pragmatic design and the real-world low-resource setting, which is exactly where these tools are most often promised and least often tested. The limits are that fourteen-day treatment failure in a mostly outpatient population is a low-event, somewhat blunt endpoint, we do not know how often the clinicians actually consulted the model, and one country's clinical officer workforce may not generalize.

Now put that next to the trial in JAMA from Zhang and colleagues, which tested something much narrower. They built an AI-based optical coherence tomography system to detect diabetic macular edema, validated it prospectively in silent mode on six hundred and three patients with diabetes, then ran a multicentre noninferiority randomized trial in two hundred and seventy-six patients with suspected macular edema referred out of a territory-wide screening programme. Under standard practice, referral is triggered by the fundus photograph alone, and about seven in ten of those referrals turn out to be false positives. Adding the AI optical coherence tomography read as a second gate dropped the false-positive referral rate to roughly a quarter — and sensitivity for referral stayed complete, with no missed cases of macular edema among patients the intervention arm declined to refer. That is a genuine specialist-capacity finding rather than an accuracy statistic. Caveats: a single health system, a modest sample, all participants ultimately got specialist review so we are measuring referral behaviour and not long-term visual outcomes, and about seven percent of scans were ungradable with a further small fraction flagged as uncertain — meaning the pathway needs a human fallback lane by design.

These two trials pull apart, and the disagreement is the useful part. One says AI assistance did not change clinical outcomes; the other says AI changed clinical behaviour dramatically. The reconciliation is almost certainly scope. The ophthalmology system had one question, one input, one decision, and a clear inefficiency to attack. The primary care model had open-ended cognition and a diffuse endpoint. What would settle it is a generative decision support trial with a targeted endpoint — antibiotic prescribing, say, or diagnostic concordance in a defined complaint — rather than a broad composite. Practically, if you are commissioning AI today, the evidence favours narrow tools inserted at a known bottleneck, and counsels patience with general-purpose copilots at the point of care.

Turning to risk prediction, where two papers make the same uncomfortable argument: our existing risk markers are much worse than we act like they are. In Nature, Obermeyer and colleagues applied deep learning to a Swedish regional dataset linking every electrocardiogram to death certificates, hunting for sudden cardiac death risk. The model isolated a small high-risk group — about two percent of the sample — with an annual sudden cardiac death rate of seven percent, higher than the rate among patients with reduced ejection fraction. And here is the finding that should stop you: the large majority of those high-risk patients, around eighty-six percent, were never flagged by ejection fraction at all. High-risk patients who happened to have a defibrillator implanted died substantially less often than expected, which hints at a mortality benefit without proving one. The model was externally validated in a United States health system, where it predicted ventricular arrhythmias, and in a Taiwanese registry, where it predicted arrhythmic cardiac arrests. Then, unusually, they paired it with a generative waveform model to visualize what the network had learned — yielding a morphology a cardiologist can actually see on the tracing, and a testable mechanistic hypothesis. The limitation is that the defibrillator benefit is observational, confounded by whoever got implanted and why; only a randomized implantation trial in ECG-flagged, normal-ejection-fraction patients settles that. But the retrospective validation across three continents makes this hard to dismiss.

The same story plays out in bone, in PLOS Medicine, where Axelsson and colleagues built and validated a hip fracture prediction tool on essentially the entire Swedish population over fifty — three and a half million people, with over a hundred and forty thousand hip fractures during follow-up. Their deep survival model, drawn from registry data with no patient visit required, discriminated two-year hip fracture risk well, and a stripped-down thirty-five-variable version performed about as well as the full one. The comparison that matters is against Sweden's current fracture liaison service approach, which screens people after a recent fracture: that strategy is barely better than a coin flip for identifying who fractures next. The model identified nearly seven times more people at genuine risk — catching roughly five in six future hip fractures versus about one in eight — at the cost of a meaningfully broader net, with specificity falling from near-perfect to about eighty percent. Notably, traditional Cox models with similar predictors performed comparably, so the win here is the data breadth, not the deep learning. The authors are candid that there is no external validation outside Sweden and no implementation study, and a Swedish registry ecosystem is not reproducible in most countries. The clinical consideration is the same for both papers: these are screening triggers, not treatment decisions, and neither has yet shown that acting on the flag improves outcomes.

Shifting to the question of what kind of model you should actually build, three papers converge on purpose-built beating general-purpose. In Nature Medicine, Kondepudi and colleagues make the case for what they call health system learning — training directly on uncurated clinical data. Their neuroimaging foundation model, built on over five million routine clinical MRI and CT volumes, outperformed frontier general models on radiologic diagnosis and, paired with open-source language models, generated reports preferred by experts with fewer hallucinated findings and fewer critical errors. That last point is the clinically relevant one: the safety gain came from domain-specific training, not from a bigger general model. This is retrospective benchmark work, not a reading-room trial, and the gap between benchmark performance and report quality in live workflow remains unmeasured.

The same pattern shows up in emergency care. In JAMA Internal Medicine, Desai and colleagues tested six widely available chatbots on simulated out-of-hospital cardiac arrest scenarios. The off-the-shelf models handled the basics reasonably — meeting roughly nine in ten minimally viable criteria — but managed only about seven in ten of the more nuanced guideline elements, things like ensuring full chest recoil. A purpose-built instructor agent hit essentially complete adherence, including on real dispatcher-assisted 911 calls, outperforming human dispatchers by a wide margin on those nuanced criteria. This is a cross-sectional proof of concept against guideline checklists, not against survival, and no real bystander was ever on the line. Which is precisely the concern that Ruhrberg Estévez and colleagues raise in PLOS Medicine, arguing that as medical AI shifts from single-task models to agents running multi-step clinical workflows, benchmarks must assess clinical reasoning, process safety and resource stewardship rather than final answers alone. A checklist score is exactly the kind of output-only metric they warn about.

Which brings us to the humans. In Nature Medicine, Xu and colleagues ran two large experiments — six hundred and twenty-three lay people and a hundred and fifty-three primary care physicians — pairing a fairness-trained dermatology AI with different explanation styles. Training for balanced performance across skin tones improved accuracy and narrowed skin-tone disparities for both groups. But large language model explanations cut differently by expertise: lay users showed strong automation bias, gaining when the model was right and losing when it erred, while experienced physicians stayed resilient and benefited either way. And showing the AI's diagnosis before the human committed produced stronger anchoring. The design is experimental with simulated cases rather than real patients, so effect sizes will not transfer directly to clinic. Still, the practical lesson is concrete: form your impression first, then open the AI, and think hard before pushing explainable AI into consumer-facing symptom checkers.

That individual-level finding sits inside two systems-level reviews. In Nature, Clusmann and colleagues systematically map security and safety hazards of large language models across the design, data, model, inference and deployment stages, classifying threats by current clinical relevance and assigning mitigation responsibility to specific stakeholders — a useful procurement checklist, though as a review it inherits the thin empirical base on real-world attacks. And in The Lancet, Dai and colleagues reframe the whole enterprise around a projected global shortfall of eleven million health professionals by 2030, arguing AI should be judged as a retention strategy — ambient documentation, coding, scheduling, inbox triage — rather than a substitute for clinicians. Their warning is the one to hold onto: AI will not fix a dysfunctional system by default, and poor implementation shifts burden rather than removing it.

If you read only one paper from this period, make it the deep learning ECG biomarker for sudden cardiac death in Nature. It is the rare AI paper that does not just predict better but hands back a visible waveform finding and a mechanistic hypothesis — turning a black box into something a cardiologist can see, teach, and test. If it holds up prospectively, it redraws who gets considered for a defibrillator.

Here is what this period adds up to. First, randomized evidence has finally arrived, and it rewards narrowness: a tightly scoped AI inserted at a known bottleneck changed practice, while an open-ended generative copilot in primary care proved safe but inert. Second, our incumbent risk markers — ejection fraction, recent fracture — are performing far worse than their entrenchment implies, and models trained on routine data find the patients they miss. Third, purpose-built beats general-purpose, in neuroimaging reports and in CPR instruction, and the safety gain comes specifically from domain training. Fourth, what remains genuinely unsettled is whether acting on any of these flags improves outcomes — every prediction paper here stops at discrimination, and none has an implementation trial. And what to watch for is the benchmarking problem: as these tools become multi-step agents, checklist and accuracy scores will stop telling us what we need to know, and the human-factors work suggests the biggest risks sit in how and when the output is shown, not in the model itself.

That's your AI in Medicine update for this period. Until next time.

This is an automated summary generated by artificial intelligence, which can make mistakes. Always review the original source materials.