Minority-Class Failure and Confidence Miscalibration in Electrocardiogram Classification: A Cross-Dataset Error Audit

Automated electrocardiogram classifiers are usually evaluated by aggregate accuracy, a number that can look strong while hiding which patients and which classes a model actually fails on. We audit this gap directly. Seven supervised classifiers, spanning linear, distance-based, kernel, tree-ensemble, boosting, and neural methods, are trained under an identical patient-level protocol on three independent, public electrocardiogram datasets: MIT-BIH (44 records, beat-level arrhythmia typing), a stratified subsample of PTB-XL (1,961 patients, diagnostic classification), and LUDB (187 records, rhythm classification). Minority classes fail at 2 to 9 times the rate of the majority class in every dataset. On the two datasets where all seven models are compared directly, AdaBoost is never confidently wrong, while gradient boosting is confidently wrong in 39.9% to 88.2% of its errors, despite similar or better accuracy. A subset of patients is misclassified by every model on LUDB, and these patients are on average nine years older than correctly classified patients wherever demographic data are available. Accuracy and confidence calibration are not the same property: two models can match on one and diverge sharply on the other, which matters because an uncertain wrong answer can be caught before harm and a confident one cannot. We argue that patient-level and class-level error auditing should accompany aggregate accuracy reporting whenever a classifier of this kind is proposed for clinical or consumer use.