What is recall?
Recall is the share of all actual positive cases that a model finds. If 50 patients on a ward develop sepsis and the early-warning model flagged 40 of them, its recall is 0.80. The formula is true positives divided by the sum of true positives and false negatives. Clinicians call the same measure sensitivity, and statisticians call it the true positive rate.
Recall measures what the model misses. It ignores false alarms entirely, so a model that flags every patient has perfect recall and no practical value. Precision covers the other side, and the F1 score combines the two. Specificity is the matching measure for the negative class: the share of healthy patients the model correctly leaves alone.
Why recall matters
Recall is the metric to optimize when a missed case costs far more than a false alarm. Undetected fraud, a missed safety defect, a late-stage diagnosis, or a critical shipment that nobody expedited all show up as false negatives. In those settings a team will accept more alerts to catch more real events, and recall is the number that tells them whether the trade worked.
Recall is also where many vendor claims fall apart. A model can look excellent on accuracy because the positive class is rare and the model mostly predicts the common outcome. Asking for recall on the positive class exposes how many real cases slip through.
Recall in practice
A community hospital runs a model that flags patients at risk of sepsis in the next six hours. Each alert sends a nurse to reassess the patient, which takes about ten minutes. The clinical team decided that missing a case was far worse than an unnecessary reassessment, so they set the confidence threshold for a recall of 0.90 and accepted that roughly two of every three alerts would be false alarms. The threshold is reviewed each quarter against the actual alert load on the floor.
What to watch for
- Always read recall alongside precision, because either one alone can be pushed to a perfect score by a useless model.
- Ask how recall was measured when positive cases are rare, since a test set with only a handful of positives gives an unreliable number.
- Decide the acceptable miss rate with the people who bear the cost of a miss, then set the threshold to deliver it.
- Re-measure recall after deployment, because the mix of cases that reach the model in production rarely matches the test set.