Answers are grounded in this paper and cite their sources.
Emergency department disposition decisions integrate measurements, narratives and contextual risk. We developed SMARTriage AI, a multi-agent system combining role-specialized language-model agents with deterministic rules, intended as safety gates, to recommend admission or discharge. In a retrospective, single-center development study of 469 valid-label adult internal medicine and geriatric encounters from one 24-h period (468 analyzed; 154 admitted), prompts and rules were refined over about 30 cycles on the same cohort, without a held-out set. Apparent accuracy was 88.2% (95% CI, 85.0–90.9%) and admission recall 96.1%. Accuracy was 97.2% with physician summary or recommendations text (n = 248) and 78.2% without it (n = 220); with the text withheld it was 65.0%. After peer review, a redesigned system (v2), evaluated on all 469 encounters under a protocol locked before the runs, discriminated worse than a single language-model agent (Setting A AUC 0.794 vs. 0.821–0.823). At the pre-specified thresholds, accuracy was 76.3% for v2 versus 67.0% for discharging everyone. The multi-agent design did not outperform a single agent, and apparent performance was optimistically biased by development on the evaluation encounters and by disposition information in physician documentation. Evaluation on unseen encounters, and external and prospective validation, are required before any conclusion about clinical utility.
No takes yet. Share an insight, caveat, or question.
Tsafrir et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: