A multi-branch vision transformer with GA-optimized focal loss for ECG signal classification.

Abstract

Background The classification of electrocardiogram (ECG) signals is a critical task in detecting cardiac arrhythmias. However, challenges such as class imbalance and the need to capture both local and global temporal patterns make this problem complex.ObjectiveIn this study, a hybrid deep learning model, called MBViT-FocalGA, is proposed for the classification of electrocardiogram ECG signals.MethodsThe proposed model combines a Multi-branch Vision Transformer (MBViT) architecture with a Focal Loss function optimized via a Genetic Algorithm (GA) called as MBViT-FocalGA. Thanks to three independent ViT branches operating at different patch sizes, the model can effectively learn both local and global timing patterns from ECG signals. Furthermore, the class imbalance problem frequently encountered in arrhythmia data is mitigated by optimizing the hyperparameters of the Focal Loss function with GA, improving the recognition of minority classes. The model's performance was evaluated on the MIT-BIH Arrhythmia Database in two different scenarios: (i) a two-class scenario combined into normal (N) and others (O), and (ii) a five-class scenario consisting of N, L, R, A, and V.ResultsExperimental results demonstrate that the MBViT-FocalGA model achieves superior classification performance with 98% overall accuracy and a high F1 score, outperforming classical deep learning models for both scenarios.ConclusionThese findings demonstrate the strong potential of transformer-based architectures and adaptive loss function optimizations in ECG classification.