Glossary

Sensitivity, specificity, and AUROC

Two ways to describe how well a clinical test or model separates sick from well — one at a single decision point, one across all of them — and why the headline number can hide the one that matters. As of July 2026.

By Jonas WeirReviewed by Jonas Weir · editorial reviewUpdated

The short version

  • Sensitivity and specificity describe a test at one chosen threshold; AUROC summarises its ranking ability across every threshold in a single number.
  • AUROC has a clean meaning: the probability a random diseased patient is scored higher than a random non-diseased one.
  • A strong AUROC does not guarantee a usable operating point — a deployed sepsis model with AUROC 0.63 still missed 67% of cases at its alert threshold.
  • AUROC ignores prevalence and the different costs of false positives and false negatives, so it can look reassuring on imbalanced clinical data while hiding poor real-world sensitivity.

Sensitivity and specificity describe a diagnostic test at one chosen threshold — the share of true cases it catches and the share of healthy people it correctly clears. AUROC, the area under the receiver operating characteristic curve, summarises how well the test ranks patients across every possible threshold in a single number. One is a snapshot at a decision point; the other is a summary of all of them. As of July 2026.

Why it matters in healthcare

The distinction decides how a result should be read. Sensitivity and specificity are the numbers a clinician lives with at the bedside, because care is delivered at one operating point: an alert either fires or it does not. AUROC is the number that dominates papers and product sheets, because it compresses performance into a single, threshold-free score. Trouble starts when the headline AUROC is treated as if it described bedside behaviour. The two answer different questions, and choosing a threshold — trading sensitivity against specificity — is a clinical decision the summary figure cannot make for you.

How it works

A ROC curve plots sensitivity against the false-positive rate as the decision threshold sweeps from strict to lenient. Its area has an elegant interpretation, first set out in 1982: "the area represents the probability that a randomly chosen diseased subject is (correctly) rated or ranked with greater suspicion than a randomly chosen non-diseased subject" 1. An AUROC of 0.5 is chance; 1.0 is perfect separation. Sensitivity and specificity, by contrast, are read off one point on that curve — the threshold actually chosen for use. Reporting both the curve and the operating point is now the expectation: STARD 2015 sets out "30 essential items that should be included in every report of a diagnostic accuracy study" 2.

Where it appears today

The gap between the two becomes vivid in deployment. A widely implemented proprietary sepsis model was found, in external validation, to have an "area under the receiver operating characteristic curve of 0.63" while it "did not identify 1709 patients with sepsis (67%)" at its alerting threshold 3. The summary figure was mediocre; the operating-point sensitivity was worse, and it was the operating point that determined which patients went unflagged.

Common misunderstandings

AUROC captures clinical usefulness. It does not. Reviewing imaging tests, methodologists note that "ROC AUC does not account for prevalence or different misclassification costs arising from false-negative and false-positive diagnoses" and that "change in ROC AUC has little direct clinical meaning for clinicians" 4. Two tests can share an AUROC yet behave very differently at the threshold a service would use.

A high AUROC survives rare outcomes. On imbalanced data — the norm for most clinical alerts, where the event is uncommon — "ROC plots in the context of imbalanced datasets can be deceptive with respect to conclusions about the reliability of classification performance" 5. Precision-recall analysis often exposes weakness that the ROC view flatters over.

Sensitivity and specificity are fixed properties of a test. They move with the threshold. Raise it to cut false alarms and sensitivity falls; lower it to catch more cases and specificity falls. Quoting either without its threshold — and without prevalence — leaves the number close to uninterpretable.

Related terms

Discrimination metrics like these say nothing about whether predicted probabilities are trustworthy — that is the job of model calibration. They shift when a model meets a new population, which is what external validation tests and what model drift erodes over time. For how these measures show up across real clinical AI evidence, see our AI in healthcare statistics.

Questions & answers

  • What is the difference between sensitivity, specificity, and AUROC?

    Sensitivity and specificity describe a test at one chosen decision threshold: sensitivity is the share of true cases it catches, specificity the share of healthy people it correctly clears. AUROC — the area under the receiver operating characteristic curve — summarises how well the test ranks patients across every possible threshold, in a single number between 0.5 and 1.

  • What does an AUROC value actually mean?

    AUROC is the probability that a randomly chosen patient with the condition receives a higher score than a randomly chosen patient without it. An AUROC of 0.5 is no better than chance; 1.0 is perfect ranking. It measures ranking ability, not whether any single threshold is clinically useful.

  • Can a model have a good AUROC but still be unsafe to use?

    Yes. AUROC ignores prevalence and the different costs of false positives and false negatives, and it says nothing about the threshold a clinician will actually act on. A model can post a respectable AUROC yet, at its chosen alert threshold, miss most true cases — so the operating point must be reported alongside the summary figure.

Sources

  1. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143(1):29-36. doi.org/10.1148/radiology.143.1.7063747
  2. Bossuyt PM, Reitsma JB, Bruns DE, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. doi.org/10.1136/bmj.h5527
  3. Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Intern Med. 2021;181(8):1065-1070. doi.org/10.1001/jamainternmed.2021.2626
  4. Halligan S, Altman DG, Mallett S. Disadvantages of using the area under the receiver operating characteristic curve to assess imaging tests: a discussion and proposal for an alternative approach. Eur Radiol. 2015;25(4):932-939. doi.org/10.1007/s00330-014-3487-0
  5. Saito T, Rehmsmeier M. The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLoS One. 2015;10(3):e0118432. doi.org/10.1371/journal.pone.0118432