Predicting spontaneous termination of atrial fibrillation based on dual path network and feature selection

Abstract

The termination mechanism of Atrial Fibrillation (AF) is still unclear. In order to avoid the potential incomprehensiveness in feature extraction via manual techniques, a Dual Path Network-SVM (DPNet-SVM) was constructed in this study for predicting the spontaneous termination of AF. The dual paths of the network extracted time and time–frequency domain features, respectively. Different sizes of depthwise separable convolution (DS-Conv) were used in both paths, which enhanced feature extraction via exploration of different sensory fields. Then four methods were used for feature selection. Finally, SVM was employed for AF termination prediction. After validating the model performance based on 1-min fragments, we conducted experiments on 10-second fragments. On the AFTDB public dataset, the accuracy, sensitivity and specificity of 95.6%, 93.3% and 91.8% were obtained, respectively. On the clinical data from Shandong Provincial Hospital (SPHD), the three metrics were 94.3%, 92.9% and 95.4%, respectively. The subsequent cross-database experiment proved the model's good generalization capability, and also substantiated its feasibility in practical applications involving wearable ECG data.