An Electronic Health Record–Integrated, Large Language Model–Powered Tool to Triage Surgical Patients
Authors' conclusion
In this prospective quality improvement study of an LLM-powered, EHR-integrated, human-in-the-loop AI system, AI-enabled screening tools accurately augmented surgical patient triage.
Study context
- Population
- Surgical cases triaged for surgical comanagement (SCM) eligibility at Stanford Health Care, September 2025 to February 2026
- Participants
- 6,193 triaged surgical cases
- Exposure
- LLM-based, electronic health record (EHR)-integrated SCM triage tool vs physician determination
- Outcome
- Sensitivity, specificity, PPV and NPV of the tool vs physician feedback as reference
Methodological detailsunblinded, human-in-the-loop, single institution, 14 of 16 SCM attending physicians participated
- Unblinded
- Human-in-the-loop
- Single institution
- 14 of 16 SCM attending physicians participated
Main finding
As reported in the abstract
Identification of patients appropriate for SCM (reference: treating physicians' determinations)
| Group | Sensitivity | Specificity |
|---|---|---|
| Tool | 0.94 (95% CI, 0.91-0.96) | 0.74 (95% CI, 0.71-0.77) |
Tool
Sensitivity0.94 (95% CI, 0.91-0.96)
Specificity0.74 (95% CI, 0.71-0.77)
Observational data: associations, not effects.
Further findings
Recommended for hospitalist consultation
- Triaged cases recommended for hospitalist consultation
- 1582 (25.5%)
False-negative cases due to LLM misclassification (post hoc medical record review)
- LLM misclassification among false-negative cases
- 2 of 19 (11%)
- Most discrepancies reflected modifiable gaps in clinical criteria, institutional workflow or physician practice variability rather than LLM misclassification
- Post hoc
Original publication
An Electronic Health Record–Integrated, Large Language Model–Powered Tool to Triage Surgical Patients. JAMA Netw Open. Published 3 August 2026. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →