Abstract
Deep learning models for automated electrocardiogram (ECG) interpretation have achieved high diagnostic accuracy, yet clinical adoption remains limited due to a lack of explainability and frequent reliance on private datasets and closed-source implementations. We present a clinically validated explainable AI (XAI) framework for multi-label ECG classification that shifts the focus from raw performance toward explainability, reproducibility, and clinical usability. Our framework is developed exclusively using publicly available resources—the PTB-XL database (21,799 ECGs across five diagnostic categories), an open-source model (Ribeiro et al. [14]), and open-source libraries. The model achieves an F1-score of 0.738 while providing gradient-based saliency maps evaluated by practicing cardiologists. We further compare time-domain and time–frequency representations (STFT and CWT). Direct time-series analysis achieves superior performance compared to transformed inputs with this model, while also requiring substantially lower computational cost. A key finding from a two-round validation process with clinicians is that explanation usefulness depends not only on the attribution method but also on presentation format: cardiologists could not reliably interpret saliency overlays until they were rendered on standard ECG grid paper used in daily clinical practice. Our contributions are: (1) a fully reproducible XAI pipeline with documented iterative clinical validation, (2) evidence that visualization format is integral to explanation quality in ECG analysis, (3) a GAM-compliant extension of the open-source ecg_plot library, and (4) a controlled analysis of signal transformation trade-offs.