Research And Application Of Multi-classifier Based Cardiovascular Disease Identification

An electrocardiogram(ECG)is the primary way to detect and prevent cardiac abnormalities.ECG contains rhythmic features of continuous heartbeats and morphological features of ECG waveforms,The nonlinearity and complexity of abnormal ECG make the waveforms of different heart diseases unique and timesensitive.The use of deep learning automatic analysis technology can autonomously excavate the potential deep-seated essential features of ECG,dig out the complex information contained in the data,effectively avoid the uncertainty of manually extracted features,and improve the accuracy of classification and recognition.Real-time analysis and diagnosis of remote ECG signals based on large-scale ECG data,to address the application requirements of two kinds of devices,namely singlelead portable wearable ECG acquisition device and standard 12-lead large ECG machine,a multi-classification analysis diagnostic model with multiple classifiers is established respectively by using a combination of multiple neural networks to solve the shortcomings of conventional neural networks that are difficult to deeply explore the essential features of ECG and improve the recognition and diagnosis accuracy.Finally,the remote ECG real-time analysis and diagnosis system is designed to realize the remote ECG auxiliary analysis and diagnosis function.The main research work includes:(1)A classification model for single-lead ECG data.For single-lead long-range time series data,RR interval features and Transformer fusion networks are established to identify and classify multiple arrhythmic heartbeats.The model utilizes a multi-headed self-attentive mechanism to model the global dependencies of ECG signals,avoiding the use of any convolutional operations and recurrent networks.The RR interval features are also incorporated into the feature vector to enrich the feature representation.Experimental results show that the RR interval features and Transformer fusion network established in this chapter have an accuracy of 99.30% in five arrhythmia heartbeat classification tasks,which has a better performance compared to other related studies.(2)A classification model for clinical multi-lead ECG data.Aiming at the structural characteristics of standard 12 lead long-term ECG signal,this paper proposes a multichannel parallel network MLCNN-Bi LSTM combining Multi-lead Convolutional Neural Networks(MLCNN)and Bi-Directional Long-Short Term Memory(Bi LSTM)to fully exploit the feature information contained in ECG.Among them,MLCNN channels are used to extract the morphological features of ECG waveforms.Compared with the traditional CNN network,the MLCNN network can accurately extract the strong correlation information on the multi-lead ECG while ignoring the irrelevant information,which is more suitable for the special structure of the multi-lead ECG.Bi LSTM channels were used to extract rhythmic features of ECG continuous heartbeat.Finally,the time-space features extracted by setting the core threshold parameters weighted to fuse multiple channels in parallel are used to explore the sensitivity of different cardiovascular diseases to morphological and rhythmic features.The experimental results showed that the identification accuracy was 87.81%,sensitivity was 86.00% and specificity was 87.76% for a variety of cardiovascular diseases with critical value attributes.(3)Establishing a remote ECG real-time analysis and diagnosis system.Aiming at the application requirements under different scenarios of single-lead and multi-lead devices,basic functions such as real-time access to ECG data,storage,user information management,remote real-time diagnosis,and doctor-patient UI interface display are completed.Finally,realize real-time analysis and diagnosis of remote ECG data.