There is a role for AI augmentation of the acute clinical encounter, and this manuscript from Science does demonstrate that potential tools are improving. However, the headlines (AI beats doctors) are of course splashier than reality — this was not a prospective ED-based study measuring EM physician performance, and in any case, diagnosis is rarely the crux of the problem we are trying to solve during the acute encounter.
This is a dataset evaluation, with both AI and physicians generating differentials offline at predefined touchpoints. No head-to-head, real-time, in-the-moment EM workflow was tested. While the study makes claims about EM, it is critical to note that the comparator physicians were internal medicine attendings — not emergency medicine. If the implication is that this informs ED triage and how precision evolves as a case unfolds, the comparator specialty matters. EM clinicians reason differently because the task is different.
The ED encounter is largely about ruling out dangerous disease (ACS, PE, SAH, sepsis, stroke), shaped by social and system factors as much as pathophysiology. While House made it seem like “the diagnosis” is the climax of medicine, in real EM practice, the rule-out — and then care coordination — is usually the work.
Physician-in-the-loop LLMs will likely help. As ED disposition becomes more granular — admit vs. transition clinic vs. hospital-at-home vs. tele-coordination — earlier, more precise diagnostic support could meaningfully change patient experience and care pathways. That said, we already have RCT-level evidence that imaging AI improves diagnosis, and even those tools remain poorly implemented secondary to infrastructure silos in healthcare delivery and reimbursement.
