Research On ECG Signal Classification Method Based On Deep Learning And Attention Mechanism

In recent years,deep learning has become an important research method in different applications including healthcare,especially in the detection of ECG signal abnormalities.However,existing deep learning-based ECG signal classification models suffer from large network model depth and over-fitting,and the accuracy of the classification models needs to be improved.To this end,this study investigates the ECG signal classification method by introducing different forms of attention mechanisms for optimisation in conjunction with deep learning,and uses the MIT-BIH dataset to train and validate them respectively.The main research components of this paper are as follows.(1)ECG signal pre-processing.Firstly,the discrete wavelet transform soft thresholding method was used to denoise the ECG signals in the MIT-BIH heart rate abnormality database;then the adaptive thresholding algorithm was used to localise the R-wave peaks of the ECG signals,and then the heartbeat segmentation was carried out based on this localisation result;finally,the ADASYN algorithm was used to process the data imbalance of the segmented heartbeat data,which enhanced the balance and reliability of the experimental data.(2)A PSTA-Net-based ECG signal classification method is proposed.This method adopts a parallel architecture model and incorporates an attention mechanism to achieve automatic classification of ECG signals.In terms of feature extraction,a spatio-temporal attention module is used to capture the anterior-posterior dependencies of ECG signal sequence data and local correlation features at different scales,respectively,to achieve the fusion of temporal and spatial features at different scales.The final classification results showed that the overall classification accuracy(OA),specificity(Spe),sensitivity(Sen),precision(Pre)and Macro-F1 of the model for the five different ECG signals in the test set were 99.50%,99.61%,96.20%,98.02% and 97.08%,respectively,achieving better classification accuracy in the field of ECG signal classification.(3)A FA-Net-based ECG signal classification method is proposed.The method avoids the currently commonly used CNN and LSTM based ECG classification models,employs a multi-headed attention mechanism,and uses location coding and dense interpolation strategies to finally obtain the feature vectors of ECG data.The final classification test of the model achieves an overall accuracy of 99.27%,a specificity of 99.33%,a sensitivity of 95.06%and a Macro-F1 of 96.29%,which shortens the network model depth and reduces the computational complexity of the model while ensuring the classification performance.The two ECG signal classification methods proposed in this paper obtained better results in terms of ECG signal classification and recognition accuracy.