Overview
Overview
This guide documents how MOHAP deployed AI to assist clinical triage decisions, reducing average wait times by 31% while improving resource allocation accuracy.
Phase 1: Data Foundation
- Audit existing patient intake forms for structured fields
- Establish minimum required fields: age, chief complaint, vitals, onset time
- Anonymise historical triage data (minimum 5,000 cases)
- Label historical outcomes: admitted / treated-and-released / referred
Phase 2: Risk Stratification Model
Train a classification model on historical data:
- High priority: requires immediate attention
- Medium priority: can wait 30-60 minutes
- Low priority: can be managed same-day
Use: Logistic regression or gradient boosting (not black-box deep learning for clinical use). Minimum accuracy: 87% on validation set before any clinical deployment.
Phase 3: Human-AI Workflow
The AI provides a recommendation, not a decision:
code
Patient Check-in → AI Risk Score (1-5)
→ Display to triage nurse with reasoning
→ Nurse reviews + may override
→ Final triage assignment recorded
→ AI learns from overrides monthly
Phase 4: Governance
- Ethics committee approval obtained
- Clinical staff trained (minimum 4 hours)
- Override rate monitored (target: < 15% overrides signals model issues)
- Monthly performance review with clinical lead
- Incident reporting protocol defined
Key Metrics
| Metric | Baseline | Target |
|---|---|---|
| Average wait (medium priority) | 52 min | < 35 min |
| Triage accuracy | 78% | > 90% |
| Nurse override rate | — | < 15% |
| Critical miss rate | — | 0% |