Evaluating diagnostic classifiers across Z-score–stratified regions: a localized ROC-based approach

Receiver operating characteristic (ROC) and the area under the curve (AUC) are widely used to evaluate diagnostic classifiers, but global performance metrics may obscure variability across regions of the prediction space. This study assessed whether Z-score–based stratification provides a localized, distribution-aware evaluation of classifier performance. Classifier predictions were standardized and partitioned into central (|Z| ≤ 1), intermediate (1 < |Z| ≤ 2), and extreme (|Z| > 2) regions. ROC, detection error trade-off (DET), calibration, and threshold analyses were performed within each region using simulated and real-world datasets. Although global AUC values suggested good overall performance, substantial regional variability was observed. Central regions showed reduced discrimination, greater class overlap, poorer calibration, and less stable thresholds, whereas extreme regions demonstrated more favorable performance characteristics. Z-score–based stratification provides a simple extension to standard evaluation methods and may help identify regional differences in classifier behavior not apparent from global summaries alone.