Three research reports landed in the first week of November, and together they are a short course in how to read medical AI claims. One was an encouraging accuracy study whose own authors urge caution. One was a large imaging study with a built-in conflict worth naming. One was a randomized trial that stopped early because it could not show patients were better off. Leaders approving AI budgets in health systems should notice which is which.
The accuracy study that says "validate us." In Nature Communications this week, researchers from NYU Langone and colleagues introduced PREVIA, an ensemble of open-source language models that predicts PTSD risk from the short trauma narratives patients give when asked what brought them to the emergency department. In 145 patients across two New York City level-1 trauma centers, the system reached an AUROC of 78.5% (±5.8%) with 95.4% sensitivity - it rarely missed future PTSD cases - but precision of only 46.8%, meaning most of its flags would be false alarms. Against a single emergency clinician scoring 30 randomly selected narratives, it performed comparably (84.4% vs 83.9% AUROC). The authors state plainly that these initial findings "warrant validation in larger, independent cohorts." That sentence is not boilerplate. It is the difference between a screening idea and a screening program.
The big imaging study, with a footnote. Also this week, City St George's, University of London announced a Lancet Digital Health study of 27,382 cardiac MRI patients in the UK and US, followed for an average of more than five years with 3,458 deaths. AI-derived ejection fraction measurements were slightly but significantly better at predicting which patients were at increased risk of death than clinician measurements, at 2-3 seconds per scan versus roughly 13 minutes of expert time. The team reports a patient-level error rate around 1% and stresses that human oversight remains essential. One detail the press coverage carried and leaders should carry too: the senior author, Dr Rhodri Davies, is the inventor of the AI tool being validated. That does not invalidate a 27,000-patient study, but it belongs in the room when the finding is weighed.
The randomized trial that found no benefit. The sobering entry came from the ALLIANCE trial, reported by ASCO AI in Oncology from an ESMO Congress 2026 presentation (Abstract LBA19) - a conference abstract, not yet a full paper. In this single-center Korean trial, oncologists treating breast and colon cancer patients could consult ChatGPT Plus during visits. At the prespecified interim analysis of 56 patients, more toxicities were recorded in the LLM arm (mean 5.4 vs 2.8), but quality-of-life scores did not improve, and enrollment was stopped. Presenting author Dr Young-Joon Kang was blunt: "on-request LLM support for the clinician was not useful in real-world everyday practice." He also flagged that without physician blinding, the team "cannot tell whether more toxicity was found or simply more was written down." The invited discussant added that only 55% of patients were assessed at the end of treatment, so even the unfavorable result carries bias risk.
The pattern for leaders. These three reports sit on different rungs of the evidence ladder. Accuracy against a benchmark or a single expert is the lowest rung: necessary, cheap, and routinely over-read. Large observational linkages to outcomes are stronger. Randomized trials measuring whether patients actually do better are the top rung, and this week's only randomized result came back negative. A vendor deck will almost always quote from the bottom rung. The question worth asking in every procurement meeting is simple: which rung is this number from, how many patients were in the denominator, and who measured it? This week offered a reminder that even good tools fail that question - and that the honest studies are the ones telling you so.
Sources
- LLM ensemble PREVIA predicts future PTSD risk from emergency department trauma narratives with expert-level accuracy - Nature Communications
- Randomized Trial Finds More Chemotherapy Toxicities Recorded With LLM Assistance but No Patient Benefit - ASCO AI in Oncology
- AI measurements of heart scans more strongly linked to patient survival than clinician measurements - City St George's, University of London
Full disclosure: Ganesh uses Instinct (instinct.ai), an AI assistant, to research, write and publish this blog series.
Top comments (0)