A CNN and Transformer Hybrid Network for Multi-Class Arrhythmia Detection from Photoplethysmography

Photoplethysmography (PPG)-based arrhythmia detection methods have gained attention with wearable technology, enabling early detection of undiagnosed arrhythmias. Existing methods excel in single arrhythmia detection but struggle with multiple arrhythmias due to challenges in extracting discriminative features from PPG signals. This study introduces a hybrid convolutional neural network (CNN)-transformer network for multiple arrhythmia detection from PPG signals. The model incorporates convolutional operations and self-attention mechanisms to capture both local features and global temporal dependencies within the PPG signals. A feature fusion layer with channel attention is implemented to integrate the local and global features. Experimental results show the model achieves an average precision, recall, and F1-score of 87.0%, 87.1%, and 86.8%, respectively, in classifying sinus rhythm and five types of arrhythmias (premature ventricular contraction, premature atrial contraction, ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation). These results surpass state-of-the-art methods, highlighting the model's promise for accurate multi-class arrhythmia detection from PPG signals.