Brugada syndrome is associated with ventricular arrhythmias and sudden cardiac death, requiring reliable detection from 12-lead ECGs. We developed an end-to-end Multi-Scale Hybrid Transformer to classify Brugada-labeled versus control ECGs directly from raw waveforms and assessed its interpretability using gradient-weighted class activation mapping (Grad-CAM). The publicly available Brugada-HUCA database comprised 69 confirmed Brugada, 7 other/atypical, and 287 control ECG. Original 12-second recordings were cropped to 10-second segments and upsampled from 100 to 500 Hz. Randomized augmentation included amplitude scaling, baseline wander, Gaussian noise, and random point dropout. Performance was evaluated using a 10-loop rotating framework with separate training, validation, and held-out test folds. Across the test folds, the model achieved a mean sensitivity of 0.671, specificity of 0.918, F1-score of 0.663, AUROC of 0.867, and accuracy of 0.871. Beat-averaged Grad-CAM highlighted clinically relevant ST-segment morphology, particularly in leads V1-V2. These findings support proof-of-concept discrimination of Brugada-labeled/atypical ECGs from controls; however, the observed sensitivity is not adequate for clinical screening, and external validation is required.