Multi-Model Comparison of Deep and Classical Architectures for ECG Diagnostic Superclass Classification on PTB-XL

Which architecture is best for sorting 12-lead ECGs into diagnostic superclasses still remains somewhat unsettled in the literature. In this study we compared six different models on PTB-XL (21,375 records) under one single patient-level split: a Mamba-style state space model (M1), a 1D residual CNN (M2), a BiLSTM (M3), a CRNN with attention (M4), a soft-voting ensemble of M2 to M4 (M5), and MiniRocket features with LightGBM (M6). Each record was labeled with the superclass of its highest-likelihood diagnostic code. We then scored the saved per-record predictions with paired patient-level bootstrap intervals, McNemar tests, per-class AUROC and AUPRC, sensitivity, specificity, and calibration. The ensemble reached the highest macro F1 (0.665, 95% CI 0.644-0.685), ahead of every single deep model by 0.020-0.041 (all paired CIs excluded zero) and ahead of M6 (0.645) by 0.021 (95% CI 0.002-0.040). We chose its members after seeing test results, so these margins are optimistic: ensembles built from validation-selected members, or from all four deep models, scored 0.656-0.657, about 0.009 lower (paired CIs included zero). M5 and M6 did not differ significantly in accuracy, and their macro AUROC was nearly identical (0.916 and 0.914). M1 trailed M2 and M3 by 0.021. Hypertrophy (HYP) was the hardest class (F1 0.33-0.41). Differences in HYP recall mostly reflected where each model's threshold sat: M6 had the lowest recall but the highest AUROC (0.871). We used only a single dataset and split, with no pretraining and no external validation, and so these results do not show that any model is ready for clinical use.