A Study On The Detection And Localization Of Myocardial Infarction By Combining Deterministic Learning And Deep Learning

Myocardial infarction(MI)is one of the most common causes of death worldwide,with 8 million deaths each year.MI is irreversible myocardial necrosis usually caused by prolonged and severe myocardium ischemia due to acute coronary artery occlusion.Early detection of MI and accurate localization of infarction are of great significance for taking effective treatment measures and saving patients’ lives.The 12-lead electrocardiogram(ECG)is a primary means of clinical diagnosis of MI.MI at different locations of the heart can be reflected in different ECG leads.However,due to various factors,many MI patients’ ECGs do not exhibit significant ischemic changes(such as ST elevation,depression,or T-wave inversion).Although there are numerous studies based on machine learning,deep learning,and deterministic learning to extract various types of static or dynamic ECG features for MI detection and localization,few studies have considered embedding clinical knowledge into the model for the detection and localization of MI.Most of the existing studies have employed a single type of feature-domain features in small-scale datasets.Therefore,this paper first utilizes clinical knowledge to guide neural network design.Then,based on deterministic learning,dynamic modeling of ECG is performed to obtain dynamic information that reflects weak changes in ECG signals.Furthermore,a multi-branch neural network is proposed to extract and fuse multiplex features such as ECG waveform features,ECG dynamic features,and ECG time-frequency features.Finally,the MI detection and localization model is trained and validated on a large-scale dataset PTB-XL.The main work of this paper is as follows:(1)Since different ECG leads reflect different sites of MI,this paper considers integrating this clinical knowledge into the design of neural network structure.The multi-branch residual neural network is designed to fuse the ECG features at different levels.Specifically,the local feature level fusion network module is used to fuse the features of each lead,and the global feature level fusion network module is utilized to fuse the features reflecting different parts of MI to obtain global features,to learn more specific and rich feature representation to improve the detection and localization accuracy.The performance of the neural network model guided by clinical knowledge is trained and validated on a large-scale dataset PTB-XL.To enhance the clinical interpretability of the model,the role of each ECG lead combination in MI localization is analyzed.(2)The dynamic change of ECG signal waveform contains abundant MI information,but few studies based on deep learning methods have considered the fusion of ECG features and dynamic features to detect and locate MI.Therefore,this paper proposes to integrate different types of ECG features.Firstly,the internal ECG dynamic information is obtained by the dynamic modeling of the ECG signal based on determination learning.Then,the multi-branch residual neural network is proposed to automatically extract and fuse ECG temporal and spatial features and its dynamic features reflecting the weak changes of ECG signals to obtain a more effective ECG representation to improve the detection and location performance of MI..(3)Compared to one-dimensional information,the time-frequency spectrum of twodimensional information is more valuable in analyzing nonlinear and non-stationary ECG signals.Therefore,this paper further integrates one-dimensional ECG signal,ECG dynamics,and two-dimensional ECG time-frequency spectrum to detect and locate MI.Firstly,based on the high-resolution Superlet transform,the ECG signal is converted into a high-resolution twodimensional time-frequency spectrum.Then,a multi-branch residual neural network is used to extract ECG spatiotemporal features,dynamic features,and time-frequency domain features respectively.Finally,these features are further integrated to construct MI detection and location models,and the model performance is verified on the large-scale PTB-XL dataset.To sum up,this study first uses clinical prior knowledge to guide and constructs a multibranch residual neural network to fuse multiple types of features.Then,based on deterministic learning,the ECG dynamic modeling is carried out to obtain the dynamic information reflecting the weak changes of the ECG signal,and the multi-branch residual neural network is proposed to extract and integrate the spatiotemporal features and dynamic features.Finally,considering the advantages of the ECG time-frequency spectrum,the multi-branch residual neural network is further improved to extract and integrate ECG spatiotemporal features,dynamic features,and ECG time-frequency features,to carry out the study of MI detection and positioning based on multi-feature fusion.The method based on multi-feature fusion further improves the model performance and achieves promising MI detection and location performance on large-scale data set PTB-XL.