Research On ECG Intelligent Analysis Based On Multi-source Feature Fusion

ECG intelligent analysis has always been the focus in the field of intelligent medicine.It is widely used in clinical ECG diagnosis,remote dynamic monitoring of ECG abnormalities,individual ECG identification,and other fields.In recent years,with the interdisciplinarity of artificial intelligence and clinical medicine,computer-assisted ECG analysis has achieved good results.Although artificial intelligence has improved the efficiency and accuracy of ECG analysis,several challenges remain.First,high-quality clinical datasets for specific diseases are scarce,and refined label of abnormal ECG signals is missing.Second,the convolutional neural network has a weak ability to express the detail features of ECG signals.The performance of ECG rhythm diagnosis,however,is still less than satisfactory because of the plurality of the meaning of ECG features.Based on the above background and problems,this dissertation proposes an ECG intelligent analysis method based on a multi-source feature fusion strategy.The automatic extraction of ECG knowledge attribute features and the fusion of morphological features are used to realize ECG-assisted diagnosis.At the same time,the neural network model is optimized and improved to further enhance the expression of ECG abnormal features,thereby improving the accuracy of ECG intelligent diagnosis.Specifically,the main research contents and contributions of this work are as follows:(1)An adaptive constrained ECG knowledge attribute feature extraction method is proposed.Previous studies extracted ECG knowledge attribute features independently from a single lead.Those methods did not consider the positional relationship between leads,resulting in a decrease in the accuracy of knowledge attribute feature extraction.To solve this problem,this dissertation proposes a knowledge attribute feature extraction algorithm to correct missing detection and false detection through adaptive constraint strategy.Firstly,key point locations(R wave,P wave,and T wave)are extracted.And traverse the single lead results and the global results to obtain the information that was undetected key points and false locations.Then,the undetected key points and false locations are automatically corrected according to the location association of ECG lead space.Finally,calculate and analyze the knowledge attribute features according to the position of the corrected key points.This algorithm can realize the refined annotation of ECG signals.In addition,a clinical ECG database named ZZU-ECG database is collected,and the performance of the ECG knowledge attribute feature extraction method is verified on the public database QTDB and ZZU-ECG database.(2)A multi-source feature fusion network model for myocardial infarction diagnosis is proposed.The myocardial infarction diagnosis relies on the knowledge attribute features and rhythm morphological features within the heartbeat.But the convolutional neural network may lose the relationship between these features as the depth increases.Compared with the previous methods that only fuse different morphological features,this method fuses the attributes of ECG knowledge.The method mainly includes two modules: the knowledge attribute features extraction module and the rhythm morphological features extraction module.The two modules can realize the mutual complementation of the two features,thereby enhancing the expression of myocardial infarction features and improving diagnostic accuracy.We collected a new database to validate the performance of the fusion strategy.The extensive experiments conducted on the PTB＿XL database and ZZU-ECG database have demonstrated the effectiveness of the proposed method.(3)A multi-branch feature fusion network model for arrhythmia diagnosis is proposed.The method relies on semantic correlation-guided feature fusion.Some ECG rhythm abnormalities will produce the same presentation of the subsequence feature.And the plurality of the meaning of ECG features will lead to a decrease in diagnostic accuracy.The proposed method adds semantic correlation analysis compared to other multi-branch network models.The purpose is to divide abnormal ECGs into different priorities according to semantic correlation and input them into corresponding branches for feature extraction and fusion.In addition,each branch also designs a feature enhancement module consisting of multi-scale dilated convolutions to integrate multi-level morphological features between ECG rhythms.Semantic correlation-guided prioritization and multi-branch network structure can effectively extract morphological features within the same priority.And fusion and enhancement module can obtain feature expression between levels,thereby improving the diagnostic accuracy of arrhythmia.Experiments conducted on the benchmark of the CPSC＿2018 database show that the feature fusion network outperforms the other diagnosis methods.(4)A remote pacing ECG recognition method based on an autoencoder model with a memory module is proposed.The current remote ECG diagnosis system only uploads the patient’s ECG,without the patient’s medical record information.And the interference caused by the mixing of pacing signal with ECG abnormal morphology makes the diagnosis more difficult for physicians.To solve this problem,we propose an autoencoder model with a memory module.The memory module is used to memorize the pacing features,the input signal is reconstructed by retrieving and matching the most similar pacing features,and the pacing ECG recognition is realized by calculating the similarity between the input signal and the reconstructed signal.We collected a paced ECG dataset to verify the effectiveness of the method.We also verified the effectiveness of the feature memory strategy on the benchmark MIT-BIH datasets.In this dissertation,a multi-scenario ECG intelligent diagnosis service system is designed to realize the application of the research.The proposed method is integrated into the system to realize ECG intelligent diagnosis service.Verified by clinical data,the proposed intelligent ECG analysis method can effectively improve the accuracy of clinical ECG diagnosis.