A Three-Class AI Model for Brugada Syndrome Detection to Improve Diagnostic Accuracy in ECG Analysis

Brugada Syndrome (BrS), a cardiac arrhythmia linked to sudden cardiac death (SCD), is diagnosed based on specific electrocardiographic (ECG) patterns, with Type 1 being diagnostic. Traditional binary classification models for BrS detection have struggled with diagnostic uncertainty, particularly in cases where Type 1 patterns are suggestive but not conclusive. Overlap with other ECG abnormalities, such as Right Bundle Branch Block (RBBB) and Non-Specific Intraventricular Conduction Delay (NIVCD), has further complicated classification. To address these challenges, a three-class system (Definitive, Borderline, non-BrS) was introduced. Initially, a binary model was adapted to ternary classification by defining probability thresholds for borderline cases, achieving approximately 85% accuracy. To fully leverage this framework, the model was modified to handle three output classes. The re-annotation process involved both the transition to a three-class system and the refinement of ground-truth labels to ensure independent classification of QRS complex and T-wave within each ECG lead. The models were evaluated through experiments on different configurations using a hold-out validation set, with the test set kept isolated for final assessment. The best model achieved 94% accuracy, with macro-average scores of 94% precision, 93% recall, and 93% F1-score on the test set. These results demonstrate that the three-class system aligns better with clinical decision-making. This study highlights the importance of integrating clinical expertise into machine learning models for complex diagnostics.