When a risk model flags a patient the clinician considers stable, the usual outcomes are that the clinician overrides it or the model is quietly distrusted. Both responses throw away information.
Three reasons for disagreement
The model knows something the clinician does not
Claims from outside the network, a filled prescription from another prescriber, an emergency department visit that never generated a note. This is the case that justifies the whole exercise.
The clinician knows something the model does not
Social context, a recent conversation, a treatment decision made last week and not yet documented. The model cannot see any of it.
The model is wrong
Miscoding, a duplicate record, a feature behaving oddly for this cohort. Genuinely common, and the only way to find it is to ask.
Capturing the override
- Offer a short structured reason, not free text alone.
- Review a sample of overrides monthly with both clinical and data staff present.
- Feed the structured reasons into the next retraining cycle.
- Report the override rate per site — large variation almost always indicates a data problem.
A model deployment without an override-review loop is not learning from the only people qualified to correct it.
