Clinically-aligned explainable AI for atrial fibrillation detection: A U-Net inspired multi-lead ECG analysis...

Deep learning offers high diagnostic accuracy but often lacks interpretability, hindering clinical adoption for Electrocardiogram (ECG) analysis. This study aims to develop a clinically interpretable, U-Net-inspired deep learning framework for Atrial Fibrillation (AFib) detection that elucidates model decision-making through morphological and rhythmic feature extraction.

We utilized a U-Net-inspired encoder-decoder architecture to analyze 12-lead ECGs as spatially structured inputs, capturing inter-lead relationships using the PTB-XL database. Model interpretability was assessed using layer-wise Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize hierarchical feature focus. Additionally, we implemented RR interval variability assessment with heart rate stratification and group-specific threshold optimization to account for physiological variations.

On PTB-XL-Improved labels, the model achieved an Area Under the Curve (AUC) of 99.15%, Sensitivity of 97.81%, Specificity of 95.63%, and Precision (PPV) of 95.72%. Performance on the original PTB-XL annotations reached an AUC of 99.78%. Grad-CAM analysis confirmed hierarchical feature extraction, correctly localizing QRS complexes in over 97.5% of cases and P-wave and rhythm irregularities in over 93.1% of cases. Independent RR-interval variability analysis achieved AUCs of 95.5% (slow heart rate), 93.2% (normal), and 91.3% (fast) across three clinically stratified heart rate groups.

This framework successfully combines high diagnostic performance with transparent, clinically aligned visual explanations. The principal contribution is not accuracy alone, but the demonstration that layer-wise explanations can be aligned with the clinical ECG reading sequence, from QRS localization to P-wave morphology. External validation remains necessary before broader clinical generalization can be claimed.