An atrial fibrillation signals analysis algorithm in line with clinical diagnostic criteria

The detection of atrial fibrillation using deep learning techniques is a hot topic in the field of signal processing. However, simply stacking modules to pursue accuracy, or compressing inputs and parameters to pursue real-time performance, leads to gambling problem between information redundancy and information loss in deep learning algorithms. At the same time, the features obtained by deep learning lack interpretability. Therefore, this study proposes a T neural network (T-Net) that integrates feature extraction, selection, and fusion. In T-Net, horizontal path extracts multi-scale information of electrocardiograms through multi-level feature reuse, feature filter embeds attention mechanism and voting algorithm internally to select information flow, and vertical path uses channel-wise point-to-point weighting to capture the nonlinear relationships of multi-scale information. Through pre-training and fine-tuning on the MIT-BIH atrial fibrillation database consisting of 23 patients, and testing on the Shandong Provincial Hospital database consisting of 252 patients, T-Net achieved accuracy, specificity, sensitivity, and F1 score of 97.95 %, 97.01 %, 98.89 %, and 97.97 %, respectively. T-Net addresses the gambling problem between information redundancy and information insufficiency, and the extracted features demonstrate good interpretability consistent with clinical diagnostic criteria, showing promising clinical application prospects.