Self-attention Neural Network Based Atrial Fibrillation Signal Extraction

With the acceleration of the aging process of the social population and the increasing work pressure of modern people,the incidence of atrial fibrillation is on the rise.The analysis of the frequency and energy of atrial fibrillation signal is helpful to study the mechanism of atrial fibrillation and track the treatment of patients,and the high-precision atrial fibrillation signal extraction algorithm can provide a more accurate basis for analysis.Therefore,it is of great significance to improve the accuracy of atrial fibrillation signal extraction algorithm.The extraction accuracy of the existing single-lead atrial fibrillation signal extraction algorithms needs to be improved,and the algorithms are sensitive to the deviation of empirical parameters and R peak detection,and can not adapt to high morphological variability data.In view of these problems,the main contributions of this paper are as follows:1)An algorithm for extracting atrial fibrillation signal based on time domain attention and DTW distance is proposed.The characteristic that DTW distance is insensitive to sampling point offset is used to reduce the adverse effect of R peak detection deviation on the extraction accuracy of atrial fibril-lation signal.Through time domain attention,the algorithm pays more attention to the extraction of atrial fibrillation signal of ECG around R peak and reduces the ventricular residual in the extracted atrial fibrillation signal.Its effectiveness is verified by simulation and real atrial fibrillation signal experiments.2)A model of atrial fibrillation signal extraction based on self-attention neural network is proposed.The problems of R peak detection steps and the setting of empirical parameters is avoided by the end-to-end training of the model.The addition of a large number of training data and the strong feature learning ability of the deep neural network makes the model more robust to high morphological variability data.The self-attention mask makes the model pay more attention to the feature learning of non-QT atrial fibrillation signals,which reduces noise interference and selects the encoder basis functions to improve the extrac-tion accuracy of the model.Through the visualization of the intermediate results of the model and the comparative experiments of various algorithms,it is verified that the model can solve the problems faced by the existing algorithms and improve the accuracy of atrial fibril-lation signal extraction.3)A model of atrial fibrillation signal extraction based on dual-channel fusion selfattention is proposed.The single-channel codec structure is extended to dual-channel,and the ventricular signal and atrial fibrillation signal are extracted respectively.The independent ventricular signal coding and decoding structure can fully learn the characteristics of the ventricular signal,and through the information interaction module in the model to help the extraction of atrial fibrillation signal.In order to increase the anti-overfitting ability of the model,the ventricular signal is coded and constrained to obtain a more robust feature space and reduce the adverse effects of sample outlier distribution.Through the visualization of intermediate results and comparative experiments,it is verified that the model further improves the robustness to high morphological variability data,and can extract atrial fibrillation signals more accurately.