Distinguishing Suspicious Brugada from Benign RBBB: Binary and Multi-Class Evaluation with a Lightweight CNN on Leads V1 and V6

Brugada Syndrome (BrS) is a genetic channelopathy creating an arrhythmogenic substrate, whereas Right Bundle Branch Block (RBBB) is a widespread and generally benign conduction delay. Despite their distinct etiology, they share a striking phenotypic mimicry on the electrocardiogram (ECG), specifically the rSR' pattern in the right precordial leads (V1-V3), which poses a complex signal classification challenge. This study evaluates a beat-level convolutional neural network trained on raw morphology from leads V1 and V6 to classify four clinically relevant patterns: Diagnostic BrS, Suspicious BrS, RBBB, and Normal tracings. The 4-class model achieves high discriminative performance for polarized patterns (AUC 0.99 for Normal and 0.96 for Diagnostic Brugada computed on a 5-fold cross-validation), while the Suspicious class remains challenging in a multiclass setting due to overlap. A focused binary classifier successfully separates Suspicious BrS from RBBB (AUC 0.97), revealing a stable decision boundary between ambiguous and benign morphologies. Gradient-based attribution confirms V1 as the dominant lead, consistent with clinical expectations, while comparative analysis against a standard V1-V3 baseline validates the critical contribution of the lateral lead V6 in distinguishing benign conduction delays. Robustness experiments under gaussian noise show stable performance within clinically realistic conditions, supporting the potential use of this lightweight model for interpretable, scalable screening of Brugada-like patterns, even in low-quality or digitized ECG recordings.