Clinically meaningful models
A clinically useful model must predict an outcome available at the intended decision time.
Loading video…
A clinically useful model must predict an outcome available at the intended decision time. External validation tests it in a different setting or population; distribution shift can reduce performance. Sensitivity and specificity quantify discrimination at a threshold, while calibration checks whether predicted risks match observed frequencies. Prediction is not a causal claim or a treatment recommendation by itself.
Worked example
If 100 patients each receive a 20% risk estimate, roughly 20 events would be expected under good calibration. Check this across risk groups and an external clinic.