Brugada syndrome is associated with an increased risk of malignant ventricular arrhythmias and sudden cardiac death, but risk stratification based on standard 12-lead ECGs remains challenging. In this study, we investigated whether graph neural networks can improve ECG-based risk prediction in Brugada syndrome by explicitly modeling spatial relationships among leads.
We analyzed a Brugada prognosis dataset including 137 patients and 411 ECG recordings. At the ECG level, 216 recordings were event-positive and 195 were event-negative. ECG signals were processed at 1000 Hz, beat-centered, and z-score normalized before training. We evaluated graph convolutional and graph attention models using two lead-connectivity topologies designed to encode inter-lead relationships. Because the task was mildly imbalanced, model performance was assessed with balanced accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve.
Graph-based models showed consistent predictive value across cross-validation. The best-performing configurations achieved balanced accuracy above 0.80 with sensitivity and specificity in a clinically meaningful range, indicating that explicit modeling of lead interactions can improve discrimination between higher-risk and lower-risk recordings. Across experiments, topology choice also influenced performance, suggesting that physiologically motivated graph design is relevant for this task.
These results support the feasibility of graph neural networks for ECG-based risk prediction in Brugada syndrome. Modeling pairwise dependencies among ECG leads may provide information not captured by conventional lead-independent approaches and could contribute to more effective noninvasive risk stratification.