Towards Explainable AI in Cardiac Diagnostic: Attention-Based Interpretation of Brugada Syndrome

Brugada Syndrome (BrS) is a genetic cardiac disorder associated with sudden cardiac death (SCD) and characterized by specific ST-segment patterns in right precordial ECG leads. Despite advances in automated electrocardiographic interpretation, conventional deep learning (DL) models remain poorly adopted in clinical settings due to their lack of transparency. In this work, it is presented an interpretable DL framework for BrS screening that leverages soft attention mechanisms to highlight diagnostically relevant ECG segments. The proposed architecture combines convolutional layers, bidirectional LSTM units, and a trainable attention module, operating on 850 ms windows centered on the S-peak. Predictions are aggregated at the lead level to reflect clinical practice. On a multi-center labeled dataset, the model achieves 93.7% window-level accuracy and 95.1% lead-level accuracy, with F1-scores of 0.95, 0.92, and 0.89 for Brugada-like, Non-Brugada, and Ambiguous classes, respectively. Attention heatmaps consistently align with ST-segment regions in true-positive Brugada cases and distinguish ambiguous from pathological inputs more robustly than threshold-based baselines. Classification errors at the lead level are reduced by over 50% compared to a legacy binary model. These findings demonstrate the feasibility of explainable AI for Brugada detection and offer a practical step toward interpretable, trust-driven ECG triage in cardiology.