Clinical Brief
Evaluating a sepsis early-warning model before it sees a patient.
A practitioner's note on the validation pipeline we ran for a hospital chain in the Gulf. Cohort selection, sub-group fairness, false-negative tolerance and the sign-off the CMO actually asked for.
Every clinical AI model that reaches the ward has passed a sign-off from someone whose signature actually matters. In this engagement, it was the CMO. What follows is the evaluation pipeline that got us to that signature.
The cohort
Six sites, thirty-eight months of retrospective data, 412,000 adult inpatient encounters. Selection excluded encounters shorter than four hours (no signal window) and encounters with incomplete vitals in the first two hours.
What the CMO asked for
- Sensitivity above 0.85 in the 65+ cohort, where onset is most missed
- Sub-group performance broken out by ward type — the model works in ICU or it does not ship
- A false-negative audit against a manual reference set from the sepsis committee
- Explanation ranking on every positive prediction, generated at inference time
I don't need the best model. I need the model my consultants will still trust in month twelve.
