An atrial fibrillation detection algorithm based on lightweight design architecture and feature fusion strategy

Background: Atrial fibrillation (AF) is one of the common types of cardiac arrhythmias, and its medical burden is continuously increasing. Wearable ECG signal analysis based on deep learning (DL) is an effective approach for screening AF. However, existing DL algorithms require extensive computational resources for AF recognition, hindering their clinical applicability. Objective: This study aims to develop a lightweight DL model to address the challenges of DL algorithms in the clinical AF recognition domain. Method: Using a distributed approach, layer-by-layer cross-guidance mechanism, and attention fusion mechanism, we designed a lightweight cross-guidance network (LCG-Net). The main path uses lightweight depth-wise separable convolutions to extract deep-level information of AF, while the auxiliary path uses standard convolutions to compensate for the weak feature expression capability of depth-wise separable convolutions. Based on the idea of mutual guidance, a layer-by-layer cross-guidance mechanism is designed to achieve information interaction and fusion between depthwise separable convolutions and standard convolutions. An attention fusion mechanism is developed based on attention mechanism and 2D convolution templates to select and precisely fuse information from different paths and different layers. Result: LCG-Net has only 39.04 K parameters and 8.16 M computations. On a clinical dataset consisting of ECG records from 252 patients, it achieved accuracy and F1 scores of 98.39 % and 98.38 %, respectively. Conclusion: The proposed LCG-Net demonstrates excellent lightweight, stability, and accuracy, holding promising prospects in the clinical diagnosis of AF.