Research On Intelligent Diagnosis Of Myocardial Infarction Based On Multi-Dimensional Feature Extraction Of Body Surface Electrocardiogram

The mortality rate of myocardial infarction(MI)among urban residents in China was62.33/100000,and the mortality rate of MI among rural residents was 78.47/100000.The mortality rate of MI is trending to rapidly ascend.In addition,the hospitalization cost of patients with MI is expensive,with an average annual growth rate of 26.89%,which is significantly higher than the growth rate of GDP.Moreover,serious time delay exists in the diagnosis and treatment of MI in China.The average time of interventional treatment of MI in China was 165 minutes,which is much longer than the gold standard of 90 minutes.Therefore,the diagnosis and treatment of MI has be the key medical and health problem in China.The body surface electrocardiogram(ECG),as the only real-time,non-invasive method for diagnosing MI in clinical,can intuitively and real-time display the electrophysiological state of the heart.Intelligent detection of body surface ECG changes and accurate diagnosis of MI are critical for MI treatment.There are still some problems in the research of intelligent diagnosis of MI recently,such as the distortion of waveform details caused by ECG signal noise reduction,the incomplete extraction of inter-lead correlation information and local spatio-temporal features.Therefore,obtaining the clean ECG signal,and comprehensively extracting the pathological features of MI from body surface ECG,have become the research focus in the intelligent diagnosis of MI.In this paper,we research on the ECG signal noise reduction,automatic detection of MI,and precise localization of MI.The main works are as follows:1.Aiming at the distortion of some waveform details in the existing research,this paper proposes an echo state network(ESN)denoising algorithm based on recursive least square(RLS)for ECG signals.According to the features of ECG signal,ESN is constructed and the RLS is used to train the network,to automatically learn the deep nonlinear and differentiated characteristics in the noisy ECG signal.Then the network can use these features to separate out clear ECG signals automatically.The experimental results show that for noisy ECG signals with 1.25 d B baseline wander,muscle artifacts,and electrode motion,the output signal-to-noise ratios are 17.13 d B,14.27 d B and 12.48 d B,respectively.This method can restore the original shape of the ECG signal while removing the complex noise,especially the detailed shape of the low-frequency characteristic wave,which provides a reliable guarantee for the subsequent feature extraction of surface ECG and intelligent diagnosis of cardiovascular diseases.2.In view of the insufficient extraction of local pathological features in the automatic detection of MI based on single-lead ECG,this paper proposes an automatic detection algorithm of MI based on the staked Sparse Autoencoder(SAE).Each SAE is constructed to obtain the optimal feature expression of the unlabeled input ECG,and then three SAEs are stacked to mine high-resolution local features of the single-lead ECG.Combined with the bagged decision tree(Tree Bagger),the accuracy,sensitivity and specificity of automatic detection of MI are 99.90%,99.98% and 99.52%,respectively,which can provide effective assistance for intelligent diagnosis of MI in clinical.3.Aiming at the insufficient extraction of inter-lead correlation information of 12-lead ECG,the precise localization algorithm for MI based on Parallel factor decomposition(PARAFAC).The local information of the 12-lead ECG is extracted by discrete wavelet transform(DWT)to form a third-order ECG tensor based on the structural characteristics of multiple leads.Then,the low-dimensional and highly discriminative lead feature of the ECG tensor is extracted based on PARAFAC.It can distinguish between 11 types of MI heartbeats and healthy heartbeats with accuracy of 99.40%,which can provide a guarantee for the precise localization of MI.4.In order to directly obtain the morphological information of the 12-lead ECG and extract the inter-lead correlation information,a precise localization algorithm of MI based on Tucker2 decomposition is proposed.The third-order ECG image tensor is constructed by using the two-dimensional ECG image and the structure characteristics of 12-lead ECG,and then the local structure characteristics of the ECG image tensor are automatically extracted based on Tucker2 decomposition.Finally,this method achieves the MI localizaiton with the accuracy of 99.67%.This study provides a strong technical support for the clinical decision-making of accurate localization of MI by multi-lead ECG.5.Aiming at the lack of lateral information of cardiac electrical activity in 12-lead ECG,the vectorcardiogram(VCG)is used to record the stereoscopic changes of the cardiac electrical activity,and a precise MI localization method based on high-order singular value decomposition(HOSVD)is proposed.The multi-scale characteristics of WT and the spatiotemporal characteristics of VCG are used to construct the VCG tensor,which containing the local information and the spatio-temporal information.The VCG tensor is compressed in the time dimension based on HOSVD,thereby removing redundant information and extracting the local spatiotemporal features.The features are fed back to the Tree Bagger,and the proposed method achieved an accuracy of 99.80% for 11 types of MI.This study provides new ideas for the intelligent diagnosis of MI.