PPG-AFNet: a lightweight and intelligible network for atrial fibrillation identification using photoplethysmography...

Atrial fibrillation (AF) is a rapid and irregular heartbeat condition that occurs due to abnormal electrical pathways in the upper chambers of the heart. Although electrocardiogram (ECG) is the gold standard for AF detection, recently several studies have been performed that demonstrate the effectiveness of AF detection using photoplethysmography (PPG) signals. Although these approaches have claimed higher accuracies, they are limited by complex preprocessing, heavier model weights, and complex architectures, and require numerous training samples to achieve higher accuracies. In this paper, we present a lightweight PPG-based AF detection technique. The proposed approach utilizes scalograms to effectively capture subtle rhythm changes in the PPG signals. These scalograms are then fed to a custom, lightweight, intelligible neural network PPG-AFNet for AF detection. The proposed technique is capable of achieving superior accuracy of 99.21% for training on as few as 960 images, making it suitable for situations where data availability is limited. Consistent performance was observed with 10-fold cross-validation across the set of experiments, outperforming all existing state-of-the-art techniques on a public MIMIC Perform dataset. The model is also explainable using interpretable saliency maps for enhanced visual interpretability. The model offers a reduced set of trainable parameters without sacrificing performance, which makes it suitable for deployment in embedded systems with limited resources. Also, the proposed model performs consistently even for a limited set of data.