Specificity

Specificity is formally defined as the proportion of true negatives correctly identified out of all actual negative cases. The formula is:

• Specificity = True Negatives / (True Negatives + False Positives)

A test with 95% specificity will correctly return a negative result for 95 out of 100 healthy patients, generating only 5 false positives. This metric is critical in screening programs where a false positive triggers invasive follow-up procedures like biopsies.

Specificity and sensitivity exist in a fundamental trade-off governed by the decision threshold. Lowering the threshold to catch more true positives (high sensitivity) inevitably increases false positives (low specificity).

• ROC curves visualize this trade-off across all possible thresholds
• The Youden Index identifies the optimal cut-point maximizing both metrics
• Clinical context dictates priority: cancer screening favors sensitivity, while confirmatory diagnostics demand high specificity

Unlike Positive Predictive Value (PPV) and Negative Predictive Value (NPV), specificity is an intrinsic characteristic of the test itself and does not change with the underlying disease prevalence in the population.

• A test with 99% specificity maintains that rate whether screening 1,000 or 1,000,000 patients
• This makes specificity a portable metric for comparing tests across different clinical settings
• PPV, by contrast, degrades dramatically when screening low-prevalence populations, even with high specificity

Low specificity cascades into significant clinical and economic consequences. Each false positive result triggers:

• Unnecessary invasive procedures: biopsies, catheterizations, or exploratory surgeries
• Psychological harm: patient anxiety during the interval between screening and confirmatory testing
• Resource waste: follow-up imaging, specialist consultations, and pathology reviews
• Diagnostic cascade: incidental findings from follow-up tests leading to further unnecessary interventions

In FDA pivotal trials for SaMD, demonstrating high specificity is often a primary endpoint to prove the AI does not increase the false positive burden on healthcare systems.

Regulatory submissions frequently require reporting specificity at a pre-specified sensitivity threshold. For example, a CADe system for mammography might be calibrated to achieve 90% sensitivity, with the corresponding specificity reported as the co-primary endpoint.

• This approach standardizes comparisons between competing diagnostic AI systems
• It prevents manufacturers from cherry-picking operating points that inflate one metric at the expense of the other
• MRMC reader studies often use this paired analysis to demonstrate that AI assistance improves specificity without sacrificing sensitivity

Specificity estimates require rigorous statistical characterization. The 95% confidence interval around a specificity measurement indicates the precision of the estimate and is heavily influenced by the number of true negative cases in the study.

• Narrow confidence intervals require large numbers of disease-free subjects
• A study with 100 healthy patients and 95% specificity has a 95% CI of approximately 88.7%–98.4%
• Sample size calculations for specificity must account for the expected prevalence in the target population to ensure adequate representation of negative cases