Lateral Leads Drive Risk Prediction in Brugada Syndrome: A Deep Learning Study

Sudden cardiac death in Brugada syndrome (BrS) poses a severe stratification dilemma, particularly for asymptomatic individuals with traditional risk markers, such as spontaneous type-1 patterns or family history. While current diagnostic protocols exclusively prioritize the right ventricular outflow tract (RVOT) via right precordial leads, emerging evidence suggests that the arrhythmogenic substrate involves a more widespread cardiomyopathy. To address this limitation, this study presents a deep learning framework based on instance-space multiple instance learning. This framework is designed to predict lifethreatening arrhythmic events in Brugada patients using basal 12-lead ECGs. The model was developed and validated on a retrospective cohort of 89 patients collected in the Piedmont Brugada registry, including 37 subjects with documented malignant arrhythmic events. On a hold-out test set, the proposed approach achieved robust prognostic performance with an area under the ROC curve of 0.74; a safety-oriented threshold optimization yielded a negative predictive value of 0.86, minimizing false negatives in screening scenarios. Beyond classification, this study provides a morphological validation of the learned features: explainability analysis using Grad-CAM revealed that the neural network consistently prioritizes lateral leads (aVL, V5, V6) over the canonical right precordial leads. Quantitative electrophysiological analysis confirmed the biological plausibility of this focus, linking the detected regions to significant conduction abnormalities, including prolonged QRS duration and altered S-wave morphology. These findings indicate that lateral depolarization abnormalities represent a critical, yet overlooked, predictor of arrhythmic risk in BrS.