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

A soft-voting ensemble of deep 1D residual CNN, BiLSTM, and CRNN-attention models (M5) achieved the highest macro F1 of 0.665 (95% CI 0.644–0.685) on PTB-XL.

• Why it matters: It addresses uncertainty about which ECG architecture best classifies 12-lead recordings into diagnostic superclasses under a consistent patient-level split.
• What they did: Six models (M1–M6) were trained on PTB-XL’s 21,375 records with superclass labels derived from each record’s highest-likelihood diagnostic code, then evaluated using patient-level bootstrap intervals, McNemar tests, AUROC/AUPRC, sensitivity/specificity, and calibration.
• The result: M5’s macro F1 exceeded each single deep model by 0.020–0.041 and exceeded M6 by 0.021 (95% CI 0.002–0.040), though margins are optimistic because ensemble members were chosen after test results.