JAMA study: AI diagnostic assistants reduce missed cardiac events in emergency departments by 27%

The JAMA chest-pain study is the kind of study the field has been asking for and rarely getting: a prospective, multicenter, outcomes-endpoint trial of an AI decision support tool measured against the outcome that matters clinically, not against a technical accuracy metric that may or may not translate into patient benefit. A 27% relative reduction in missed 30-day major adverse cardiac events is a large effect, and the “without a significant increase in admissions” finding is the answer to the usual counter-hypothesis — that AI-driven triage would push borderline patients into inpatient beds and cost more than it saved.

Two caveats deserve attention alongside the top-line result. First, the 42-site sample is a self-selected group of emergency departments that opted into the study and had the operational maturity to integrate a new decision support tool into their workflow. Whether the same effect size shows up in emergency departments without that operational maturity is the generalizability question, and the study cannot answer it. Second, the “usual care” comparator in a study like this is heterogeneous — some sites likely had informal risk-stratification workflows already, and others did not. The 27% effect is the average across that heterogeneity, and the effect at any individual site could be larger or smaller.

For emergency medicine leaders, the study’s most useful contribution is that it gives them permission to make the operational investment. AI chest-pain triage tools have existed for several years, but the case for deploying them has rested on retrospective validation studies that clinical leaders have justifiably discounted. A prospective outcomes study at this scale, published in JAMA, changes the internal politics of the deployment decision.

For the AI-medical-device regulatory conversation, the study is a data point in favor of the argument that clinical outcome studies are feasible, affordable, and interpretable for AI-enabled devices — and that when they are done, they can generate the kind of evidence that changes clinical practice. Expect this study to be cited in the ongoing debate about what the FDA should require of AI-enabled device manufacturers.

Related coverage: clinical decision support topic · emergency AI topic.