Emergency departments and primary care services face chronic demand that exceeds capacity. In many European healthcare systems, emergency department wait times exceed 4 hours for non-urgent cases, while genuinely urgent patients may not be identified quickly enough. AI triage systems provide consistent, evidence-based initial assessments that help clinicians prioritize patients and allocate resources effectively.

The Problem

Manual triage depends on the experience and judgment of the triaging clinician, typically a nurse using a standardized framework (Manchester Triage System, ESI, or local equivalents). Triage accuracy varies with clinician experience, cognitive load, and time pressure. Under-triage (assigning a lower acuity than warranted) creates patient safety risks. Over-triage (assigning a higher acuity than necessary) diverts resources from genuinely urgent patients.

Primary care faces a related challenge: patient demand for appointments exceeds availability, and telephone triage to determine urgency is time-consuming and inconsistent. Patients with urgent needs may wait days for an appointment while routine requests consume same-day slots.

AI Approach

Symptom assessment - Bedrock powers a structured symptom collection interface (chatbot or guided questionnaire) that gathers relevant clinical history before the patient sees a clinician. The system asks adaptive questions based on presenting complaints, using clinical reasoning to identify red flag symptoms that indicate high acuity. The output is a structured symptom profile that saves clinician time and ensures no critical questions are missed.

Acuity scoring - SageMaker models predict clinical acuity (urgency of required care) from the symptom profile, vital signs (where available from triage assessment), demographic factors, and medical history. The model is trained on historical triage data linked to outcomes: hospital admission, intensive care admission, time-sensitive interventions, and deterioration events.

Resource recommendation - Based on the acuity score and likely clinical pathway, the system recommends the appropriate resource: emergency department, urgent care, same-day primary care, routine appointment, or self-care with safety-netting advice. This demand routing reduces inappropriate ED attendance and ensures primary care capacity is directed to the most clinically appropriate patients.

Deterioration prediction - For patients already in the ED waiting area, continuous monitoring of vital signs and wait time triggers re-triage alerts when predicted deterioration risk exceeds thresholds. DynamoDB stores patient state for real-time monitoring across the department.

Architecture

Patient-facing symptom collection runs via a web interface or messaging integration, powered by Bedrock for conversational interaction. Structured symptom data flows to SageMaker for acuity scoring. Results are integrated with the hospital’s electronic health record (EHR) and patient flow management system. DynamoDB maintains real-time patient state for ED monitoring. Clinical dashboards display prioritized patient lists with AI-generated summaries.

Key Considerations

Clinical validation - AI triage must be validated against clinical outcomes, not just against existing triage decisions (which may themselves be inaccurate). Prospective validation studies comparing AI triage with clinician triage are essential before deployment.

Safety-first design - The system should err toward over-triage rather than under-triage. Missing a high-acuity patient has far greater consequences than unnecessary urgency. Sensitivity for high-acuity conditions should exceed 95%.

Clinical override - Clinicians must be able to override AI triage decisions without friction. The AI informs rather than determines the clinical decision. Override patterns should be monitored to identify systematic model weaknesses.

Cross-referencing - Patient triage connects to medical imaging (imaging prioritization), appointment scheduling (demand routing), and health monitoring (remote triage for monitored patients).

Next Steps

Retrospectively validate an acuity prediction model against historical ED data, measuring sensitivity for high-acuity patients. If sensitivity exceeds 95% for the highest acuity levels, pilot the symptom assessment tool with a small patient cohort, comparing AI-informed triage against standard triage accuracy and clinician satisfaction.