Beyond the RVOT: Prognostic ECG Feature Discovery in Brugada Syndrome via Transfer Learning and Lead Ablation

Arrhythmic risk stratification in Brugada Syndrome (BrS) patients is crucial for guiding clinical management, yet challenging due to the scarcity of high-risk events and the morphological variability of 12-lead electrocardiograms (ECGs). This study introduces a deep learning (DL) framework based on a Lightweight Residual Network (L-ResNet), which employs a compact convolutional neural network initially pre-trained on the Brugada pattern diagnostic task (with diagnostic accuracy of 96.5% on a three-class task) and subsequently fine-tuned using a retrospective Brugada cohort comprising 63 patients with documented events and 74 asymptomatic individuals. The method operates at beat level and aggregates predictions via both direct and hierarchical schemes. Two aggregation strategies were compared: a computationally efficient Beat-to-Patient pipeline and a multi-level Beat-to-ECG-to-Patient approach. Results demonstrate comparable performance (AUC = 0.757, Accuracy = 0.708). A post-hoc interpretability module based on lead ablation and integrated gradients confirm the dominant role of leads V1 and V2, but unexpectedly highlights lateral leads such as aVL and V5 as critical predictors of risk. The findings underscore the potential of interpretable DL to not only predict clinical outcomes but also to uncover previously underappreciated pathophysiological markers, supporting more refined risk stratification strategies in BrS.