Source-Only Transportability of Engineered ECG Features for Healthy-Versus-Myocardial Infarction Classification

Background/Objectives: Electrocardiogram (ECG)-based myocardial infarction (MI) classifiers may achieve high internal validation performance but show reduced performance when applied to data from another source. The task is a controlled binary healthy-versus-MI benchmark and is not intended to represent real-world chest-pain triage or autonomous clinical deployment. This study evaluated the source-only transportability of engineered 12-lead ECG feature families for binary healthy-versus-MI classification across a cardiologist-annotated hospital dataset and PTB-XL. Methods: The hospital dataset contained 1749 usable recordings from 1434 patients after excluding 206 broken-data records, with 1550 Healthy and 199 MI recordings. The matched PTB-XL binary subset contained 14,982 recordings from 13,436 patients, with 9513 Healthy and 5469 MI recordings. Eleven engineered feature families and five classifier families were compared under preprocessing, patient-aware splitting, source-validation hyperparameter and threshold selection, and bootstrap uncertainty estimation. The reported leading rows are the highest observed configurations in a prespecified benchmark grid, not locked clinical models. Results: Internal performance was higher than strict source-only transfer performance. In the hospital dataset, fiducial interval descriptors with Extra Trees reached balanced accuracy 0.775 and receiver operating characteristic area under the curve (ROC-AUC) 0.855. In PTB-XL, a broad hybrid feature bank with ST-segment information and XGBoost reached a balanced accuracy of 0.898 and ROC-AUC of 0.965. Strict source-only transfer was weaker and asymmetric: the highest observed balanced accuracy was 0.580 for hospital-to-PTB-XL transfer and 0.632 for PTB-XL-to-hospital transfer. Ranking transportability and operating-threshold transportability diverged, most notably for hospital-to-PTB-XL transfer, where ROC-AUC was 0.774 but sensitivity at the source-selected threshold was only 0.164. A secondary target-threshold analysis improved balanced accuracy to 0.682 and 0.640, respectively, but this used target labels only to re-select the operating threshold and was not a strict source-only result. Conclusions: The findings indicate a transportability gap: PTB-XL-to-hospital transfer was more balanced than hospital-to-PTB-XL transfer, but neither direction achieved performance comparable to internal validation. The source-only operating-point results are not acceptable for clinical MI screening or decision support without additional calibration, target-setting validation, and prospective assessment.