Interpretable ECG Arrhythmia Classification via 1D-CNN and Grad-CAM: Evaluating Lead-Wise Attention and Cross-Institutional Generalization on PTB-XL

Deep learning models for automated ECG interpretation are typically evaluated only in-domain, leaving open whether they generalize to recordings from other institutions. We present ECGNet, a compact 1D convolutional network (~47,000 parameters) trained on PTB-XL for multi-label classification across five diagnostic superclasses (NORM, MI, STTC, CD, HYP), reaching macro-AUC 0.920 ± 0.002, within 0.008 of published results from architectures with orders of magnitude more parameters. We evaluate cross-institutional generalization on two external, unseen datasets (Georgia, Chapman-Shaoxing). Performance degrades substantially without adaptation (macro-AUC 0.670 and 0.735, averaged across three independent training runs). Recomputing Batch Normalization statistics from a small unlabeled target batch (AdaBN), without retraining or labels, partially recovers performance (0.753, 0.828), closing roughly a third to half of the gap but not eliminating it. Critically, this recovery is uneven and dataset-dependent: across three independent runs, AdaBN's effect on HYP is consistently small on Georgia but consistently large on Chapman-Shaoxing (non-overlapping ranges across all three runs), despite identical technique and model; a similar directional pattern for MI does not replicate as reliably across runs. Grad-CAM analysis shows attention broadly consistent with diagnostic criteria and surfaces a specific, explicable failure mode in the HYP class. We discuss implications for deploying compact ECG classifiers across heterogeneous clinical settings and leave the dataset-dependent AdaBN pattern as an open question.