Research On Ecg Signal Processing Techniques Based On Sparse Features And Machine Learning

With the increase of life pressure and population aging,the number of patients with cardiovascular diseases increases year by year and the mortality rate remains high.Cardiovascular disease has become a high-risk disease affecting people's health,which not only causes great physical and mental pain for patients,but also brings heavy economic burden to our society.As a reflection of cardiac electrical activity,ECG?Electrocardiogram?can record the function and state of the heart non-invasively and in real time.It is the main basis for the diagnosis of cardiovascular diseases and plays an important role in clinical practice.However,the ECG signal is an extremely weak electrophysiological signal,which is easily affected by all kinds of external noise and interference in the process of acquisition.In order to ensure the quality of ECG signal acquisition and to facilitate the diagnosis of diseases,it needs to be processed for noise reduction and feature extraction.In this context,this paper follows the research idea of"theoretical analysis?signal noise reduction?waveform detection".This paper starts from the essential characteristics of ECG signal and bases on the actual needs of diagnosis and treatment of cardiovascular diseases.Aiming at the problems that traditional denoising methods lose ECG signal details and feature extraction methods rely excessively on manual feature selection.The denoising technology of ECG signal based on sparse signal processing and the feature extraction technology of ECG signal based on machine learning are studied.The main innovations in this paper are reflected in the following four aspects:?1?Aiming at the problem of ECG signal noise interference,this paper uses the sparseness of the ECG signal to model the signal as the sum of low-pass component,sparse component and noise component.An ECG signal denoising model combining low-pass filtering and sparse recovery is proposed.Verification by real ECG data,the results prove the correctness of the theoretical model.?2?Aiming at the problem of underestimation of the real ECG signal waveform caused by the traditional sparse recovery algorithm mostly using the ₁-norm penalty term.The GMC?Generalized Minimax Concave?function with nonconvex characteristics is introduced as the penalty term in this paper,which can greatly improve the extraction capacity of sparse components in the denoising model and effectively solve the problem of underestimation of signal waveform.The algorithm is verified by the MIT-BIH ECG database and compared with multiple noise denoising algorithms.The results show that the sparse ECG signal denoising algorithm based on the GMC penalty term can save key features in ECG signal while achieving excellent noise reduction results.?3?Aiming at the problem of traditional QRS complex detection methods mainly relying on a priori expert knowledge to select fixed features and parameters.And the shortcomings of ignoring the correlation information between leads and using only single-lead ECG signals for detection,a convolutional neural network model with multi-lead data fusion is proposed to automatically detect QRS complex.The model supports multi-lead ECG data input without preprocessing steps.And two kinds of convolution layers with different scales are designed in the convolutional neural network,one is to extract the time interval and variation information between the current wave group and the adjacent wave group,and the other is to extract the correlation information between different leads.The multi-lead ECG signal is fully utilized,and the detection accuracy of the QRS complex is effectively improved.?4?Aiming at the problem of incorrect detection and missed detection of the proposed QRS model detection results.In this paper,a backtracking module is designed for the output of the model to reduce the error detections and improve the waveform detection accuracy.Experimental results show that the proposed QRS cpomplex detection algorithm based on convolutional neural network achieves 99.74%sensitivity and 0.294%error detection rate in MIT-BIH ECG database with only two leads.In the INCART database with 12 leads,the sensitivity was 99.96%and the error detection rate was only 0.047%.It is proved that the algorithm has better performance and better robustness under multi-lead data.