Sensitivity and specificity — the two error rates

Two questions define a classifier’s errors. Of the patients who have the condition, how many does it catch? That is sensitivity — misses are its failure. Of the patients who don’t, how many does it correctly ignore? That is specificity — false alarms are its failure.

The pair matters to governance because the threshold is a policy decision wearing a technical costume. Lower the alert threshold and sensitivity rises while specificity falls: fewer missed cases, more alarms, more alert fatigue, and eventually a human-in-the-loop control eroded by sheer volume. That trade-off belongs to clinical leadership, in a minute, not to a default setting nobody chose. It is also why a monitoring plan tracks the pair at the operating point alongside AUROC — AUROC can hold steady while the threshold’s real-world behaviour shifts under changed prevalence.