Machine learning4 min read

How models are evaluated

Accuracy is only the beginning — sensitivity, specificity, robustness and calibration each tell part of the story.

“Accuracy” sounds simple, but in medicine it can mislead. If 99% of images are normal, a model that calls everything normal is 99% accurate — and completely useless. That is why medical evaluation uses richer metrics.

Sensitivity asks: of the cancers present, how many did the model find? Specificity asks: of the healthy women, how many were correctly reassured? AUC summarises the trade-off. Calibration asks whether predicted risks match reality. Each metric answers a different clinical question.

Metrics that matter

  • Sensitivity — cancers not missed
  • Specificity — healthy women not alarmed
  • AUC — overall discrimination power
  • Calibration — do risk estimates match outcomes

In brief

Raw accuracy can mislead on imbalanced medical data.
Sensitivity and specificity answer different questions.
Calibration says whether risk scores mean anything.

Keep learning

How machine learning learnsAI HubEnsembles: many models, one stronger decisionAI HubPerformance and benchmarksAI HubClinical validation and evidenceAI Hub