Machine-learning classification of myocardial infarction and ST/T-change ECG phenotypes across complementary evaluation settings

The primary PTB-XL outcome was an operational ECG phenotype defined by myocardial infarction and/or ST/T-change diagnostic superclass labels, rather than adjudicated active ischemia. Public ECG datasets were harmonized to 10-s, 12-lead, 100-Hz inputs. Patient-disjoint PTB-XL partitions were used to compare an age/sex baseline, a downsampled-waveform logistic-regression model, a compact residual network, MiniRocket–ridge, and a fusion model. In the primary held-out test set, AUROC/AUPRC were 0.649/0.654 for the linear waveform model, 0.964/0.967 for the residual network, and 0.953/0.960 for MiniRocket. Exploratory subgroup AUROCs for the unchanged composite-trained MiniRocket model were 0.923 for MI-only, 0.969 for ST/T-change-only, and 0.998 for combined MI + ST/T-change ECGs. In the PTB Diagnostic ECG Database, MiniRocket achieved an AUROC of 0.936 (95% CI, 0.899–0.965) and AUPRC of 0.987 (0.977–0.993), versus 0.924 (0.888–0.954) and 0.983 (0.972–0.991) for the residual network; paired analyses showed no clear difference in external discrimination. The 0.5 threshold was descriptive, while validation-derived thresholds were applied unchanged externally. In STAFF III, model probability increased during coronary occlusion in 51/80 paired comparisons, with changes dependent on baseline definition. Overall, the models classified dataset-defined ECG phenotypes, while the STAFF III findings support only an exploratory, baseline-dependent association with controlled occlusion. Clinical evaluation requires representative cohorts with adjudicated outcomes.