Research On ECG Analysis Model Based On Local And Temporal Features

With the increasing number of patients with myocardial infarction,many patients cannot be diagnosed in a timely and accurate manner.However,the electrocardiogram(ECG)analysis algorithm based on deep learning can realize rapid and accurate analysis and diagnosis of myocardial infarction,and alleviate the problem of difficulty in seeing a doctor.The automatic analysis algorithm of myocardial infarction still has limitations,mainly reflected in the inaccurate positioning of the ECG waveform under the influence of noise and pathological factors,which leads to a decrease in the accuracy of disease detection;there are differences in sample segmentation under the influence of disease,heart rate,and different individuals,Affecting feature extraction leads to lower accuracy of disease detection;poor generalization of tests across different patients.Aiming at the problems mentioned above in ECG analysis,this paper proposes different solutions and conducts model research on waveform positioning and ECG characterization respectively.The main research contents and contributions of this paper are as follows:(1)For the problem of inaccurate waveform positioning,a local mask attention U-Net segmentation model(Local Mask Attention and U-Net,LMAU-Net)is proposed to improve the accuracy of ECG QRS wave positioning.This method takes a single-lead ECG as input,extracts the waveform features and semantic features of ECG through the U-Net model,introduces a local mask attention mechanism to focus on changes in local waveforms,and reweights the original features to suppress noise and non-QRS wave features.Experiments were conducted on seven public ECG datasets to verify the generalization,real-time,and noise robustness of the model,and an average detection accuracy of 99.71% was obtained.The experiments show that LMAU-Net is a fast and accurate QRS wave localization model,which provides a strong guarantee for ECG analysis.(2)In the face of the problem of ECG sample segmentation differences,a myocardial infarction detection model(Multi-branch convolutional neural network and Transformer,MBC-former)is proposed that combines multi-branch convolutional neural network and Transformer.Among them,the multi-branch network can separately characterize the independent local waveforms of the 12-lead ECG to highlight the independent characteristics of each lead.The Transformer module analyzes the relevant features of the waveforms before and after the ECG from the time dimension,and mines the timing correlation of different waveforms.The model alleviates the impact of sample segmentation differences on myocardial infarction detection through the complementarity of multiple branches and the global dependence of features.Experiments were carried out on the public PTB Diagnostic ECG Database(PTB),and the average accuracy,average sensitivity,and average specificity were99.98%,99.97%,and 99.98%,respectively,using 5-fold cross-validation detection.It is proved that the method has high detection performance,which can provide a reference for the system design of clinical monitoring disease analysis.(3)Aiming at the problem of poor detection generalization among patients,a myocardial infarction detection model based on multi-level feature fusion is proposed,which includes three modules: scanning module,reading module,and thinking module(Scanning,Reading,and Thinking,SRT-Net).The scanning module is used to extract the independent waveform features of the superficial layer of each lead;the reading module is used to focus on the middle layer features of the cross-leads;the thinking module is used to integrate the global features of ECG,analyze the time and shape relationship between different waveforms,and finally use The classification function classifies myocardial infarction.The multi-level structure is designed with reference to the process of doctors diagnosing diseases,which is more conducive to capturing the general characteristics of ECG,alleviating the impact of individual differences among patients on classification.Using 5-fold cross-validation on the PTB dataset achieved an average accuracy of 97.15%,an average sensitivity of 98.28%,an average positive predictive value of 98.26%,and an average F1 score of 98.25%.Compared with other methods,the SRT-Net model obtains the best detection performance and provides a new solution for clinical auxiliary diagnosis.