Anaes · Measurement & monitoring physics

Diagnostic test performance: sensitivity, specificity, predictive values, likelihood ratios and ROC

Also known as Sensitivity and specificity · Predictive values · Likelihood ratios · ROC curves · Bayes' theorem · Screening versus diagnostic testing · STARD guidelines · Diagnostic test accuracy

Every anaesthetic decision rests on a test — a troponin, a D-dimer, a lactate, a FAST scan, a Mallampati score, an AI flag on a CT — and the framework that tells you whether to trust that test is diagnostic test performance. The model rests on seven exam-critical ideas. First, every test result can be laid out in a 2x2 CONTINGENCY TABLE of disease present or absent against test positive or negative, giving four cells — TRUE POSITIVES (TP), FALSE POSITIVES (FP), FALSE NEGATIVES (FN) and TRUE NEGATIVES (TN) — from which every metric is derived. Second, SENSITIVITY (Sn) is the TRUE POSITIVE RATE, TP divided by (TP plus FN), the probability the test is positive GIVEN disease; because it is computed only in those who have disease it is a fixed property of the test, independent of prevalence, and a highly sensitive test used to RULE OUT disease gives the mnemonic SnNout. Third, SPECIFICITY (Sp) is the TRUE NEGATIVE RATE, TN divided by (TN plus FP), the probability the test is negative GIVEN no disease, also prevalence-independent, and a highly specific test used to RULE IN disease gives the mnemonic SpPin. Fourth, the PREDICTIVE VALUES answer the clinical question: POSITIVE PREDICTIVE VALUE is TP divided by (TP plus FP), the probability of disease GIVEN a positive test, and NEGATIVE PREDICTIVE VALUE is TN divided by (TN plus FN), the probability of no disease GIVEN a negative test — and unlike sensitivity and specificity the predictive values DEPEND ON PREVALENCE, so a test that performs well in a high-risk ICU population may generate mostly false positives in a low-risk screening clinic. Fifth, BAYES' THEOREM links the pre-test probability (the prevalence) to the post-test probability through the LIKELIHOOD RATIO: post-test odds equals pre-test odds times the likelihood ratio. Sixth, the LIKELIHOOD RATIOS combine sensitivity and specificity into a single prevalence-independent number — LR+ equals sensitivity divided by (1 minus specificity) and LR- equals (1 minus sensitivity) divided by specificity; an LR+ greater than 10 or an LR- less than 0.1 is very useful, and because LRs do not depend on prevalence they travel with the test from the validation population to your patient. Seventh, the RECEIVER OPERATING CHARACTERISTIC (ROC) curve plots sensitivity against (1 minus specificity) for every possible threshold and the AREA UNDER THE CURVE (AUC) summarises overall discrimination — 0.5 is no better than chance, 1.0 is perfect, and an AUC greater than 0.9 is excellent; the chosen threshold trades sensitivity against specificity, so a SCREENING threshold is set for high sensitivity (do not miss cases — D-dimer for PE, lactate for sepsis, HIV screening) while a CONFIRMATORY threshold is set for high specificity (do not raise false alarms — biopsy, coronary angiography). New tests are evaluated against a REFERENCE STANDARD under the STARD reporting guidelines, guarding against SPECTRUM BIAS (the study population does not match the clinical one) and VERIFICATION BIAS (only positive tests get the gold standard). Built on the Altman-Bland diagnostic-tests trilogy (sensitivity/specificity, predictive values, ROC plots), the Hanley-McNeil AUC paper, the Jaeschke-Guyatt-Sackett Users' guides (validity and results), the Ransohoff-Feinstein spectrum-bias paper, the Begg-Greenes verification-bias paper and the STARD 2015 statement, with applied examples from the iridocorneal-angle AI study (Rubegni 2026), the gadoxetate-MRI PSC study (Sandrasegaran 2026), the cerebral-autoregulation CAI study (Albanese 2026), the glaucoma-referral AI study (Lima-Cabrita 2026) and the AI-CT MACE cohort (Xu 2026).