ECG-XAI: Attention-ResNet1D for Multi-Label Cardiac Diagnosis with Explainability

This paper proposes a deep learning framework for multi-label electrocardiogram (ECG) classification that explicitly addresses evaluation reliability and clinical interpretability. Using the large-scale PTBXL dataset, we employ a subject-aware cross-validation protocol and a robust preprocessing pipeline to mitigate class imbalance and patient-level data leakage. The proposed ResNet1D-SE architecture integrates residual learning with channel-wise feature re-weighting, enabling the model to prioritize diagnostically relevant temporal patterns in 12-lead ECG signals. The framework achieves a mean AUC-ROC of$0.9293 \pm 0.0046$and a mean F1-score of$0.7863 \pm 0.0174$, outperforming baseline models with low performance variability. To enhance clinical transparency, post-hoc explainable artificial intelligence techniques are applied to identify ECG regions contributing to model predictions, demonstrating consistency with established electrophysiological characteristics. The results support the development of accurate and interpretable deep learning systems for automated multi-label ECG diagnosis.