Every risk model demo looks good. Feed it two years of claims, hold out the third, and the AUC lands somewhere respectable. The trouble starts when the model meets a population that does not resemble the one it learned from — and that happens more often than the validation slide suggests.
The denominator decides the answer
Most published risk models are trained on patients who were continuously enrolled and continuously engaged. That is a convenient cohort and an unrepresentative one. Patients who disengage, move plans, or only appear in the data when something goes badly wrong are precisely the ones a care management programme exists to find.
Exclude them during training and the model learns to predict utilisation among people who were already being watched. It will do that well. It will also miss the patient who has not been seen in fourteen months.
Three failure modes worth testing for
Label leakage through billing timing
Claims arrive late and are adjudicated later. If the feature window and the outcome window overlap even slightly, the model learns the future from the paperwork rather than the clinical picture. The symptom is suspiciously high performance on retrospective data and a sharp drop in production.
Access confounded as risk
Utilisation measures contact with the system, not need. A model trained on cost will reliably rank patients with good access above patients with greater clinical severity and worse access. Rank by predicted deterioration instead, and validate against clinical endpoints.
Stability across sites
Coding intensity varies between practices far more than disease prevalence does. A model that is not checked site by site will quietly recommend more intervention wherever documentation is most thorough.
What a defensible validation looks like
- Hold out by time and by site, not by random row.
- Report calibration, not only discrimination — ranking is useless if the probabilities are wrong.
- Stratify performance by age, payer, language and visit frequency.
- Track how many flagged patients a team can actually act on; a model that surfaces 900 names for a team of four has not helped.
None of this is exotic. It is simply the difference between a model that scores well and a model that changes what a care team does on Monday morning.
