AI Medical Triage: Life-or-Death Limits

AI triage assistants can flag deteriorating vitals quickly but the ESI level and treatment priority remain a clinician's call.

What AI does well here

• Detect early sepsis signals
• Summarize prior ED visits
• Highlight medication interactions

What AI cannot do

• Set a final ESI level
• Allocate scarce resources during MCI
• Replace bedside clinical judgment

Why AI triage stays advisory in emergency medicine

Emergency Severity Index (ESI) scores determine which patients get seen first in an emergency department and which wait. The stakes are literal life and death: a patient incorrectly downgraded from ESI 2 to ESI 3 may wait an hour before evaluation and deteriorate during that window. AI triage assistants have demonstrated genuine value in specific functions: detecting early sepsis signals from vitals that humans might miss in a busy triage bay, summarizing prior ED visits to surface relevant history, and flagging drug-drug interactions for patients on complex regimens. These are augmentation functions. The key constraint is that the ESI determination itself — the classification that governs care priority — remains a licensed clinician's call. This is not merely a policy choice. Clinicians apply judgment that extends beyond the data: patient appearance, the quality of pain reporting, subtle behavioral cues, and contextual factors that do not fit into structured fields. AI systems trained on historical ESI data will also replicate historical triage biases that disadvantaged women presenting with cardiac symptoms and patients with chronic pain conditions. Any AI triage integration must include a one-click override mechanism that makes the AI's evidence visible and the human's override path unambiguous.

• AI triage value is in augmentation: sepsis signals, prior-visit summaries, medication interactions
• ESI level and care priority remain a licensed clinician's determination
• Build explicit override UI — the AI's evidence must be visible and one-click dismissible
• Audit AI triage outputs for racial, gender, and chronic-pain bias before deployment