Triage at primary care: red flags the AI must surface

Three ways to see it

1. 1

   Red flag categories that any primary care triage system must handle: chest pain with cardiac features, stroke warning signs, severe dehydration in children, ectopic pregnancy presentation, suspected meningitis, severe respiratory distress, and uncontrolled bleeding. The model lists these on top of the screen, with a single tap to confirm or reject. The medical officer remains the final authority.

2. 2

   Two false-positive policies are common and both have costs. Liberal flagging catches more genuine cases but overwhelms referral. Conservative flagging stays manageable but misses cases. For low resource settings the right answer often lies in a tiered approach: high specificity for full referral, plus a softer flag for in-clinic reassessment by the medical officer within twenty minutes. Document the calibration explicitly so it can be revisited.

3. 3

   Governance. A triage model must be approved by the clinical lead of the facility, monitored monthly for missed cases, and audited quarterly for fairness across gender, age, and rural-urban patient mix. Pakistani facilities should keep a simple incident log: every flagged case, what the doctor decided, and the outcome on follow-up. This log feeds retraining and is the first thing a regulator or donor will ask to see.

The why-tree

Try this with Claude

Implementation steps. One: write the red-flag list with two clinical experts. Two: implement on a single BHU for eight weeks in shadow mode. Three: compare AI flags to senior physician retrospective review weekly. Four: only enable visible flags after the false-negative rate on critical conditions is acceptable to the clinical lead. Five: train all BHU staff on how to override the AI, not how to follow it.