Research On Arrhythmia And Myocardial Infarction Diagnosis Algorithm Based On Deep Learning

With the continuous development of human society,the harmful effects of cardiovascular diseases on human health have become increasingly prominent.Electrocardiogram(ECG)signals are a common tool for diagnosing cardiovascular disease,and its manual analysis requires a lot of time and effort from doctors,and limits the application of daily situations such as family and community monitoring.Recently,with the development of intelligence healthcare technology,the use of deep learning algorithm models based on ECG signals to automatically diagnose cardiovascular diseases has emerged in an endless stream.Through the deployment of the automatic diagnosis algorithm,long-term and high-efficiency analysis of the ECG signal can be realized,the burden on the doctor can be alleviated,and the diagnosis result can be given in time to improve the prognosis of the patient.Therefore,this paper has deeply studied the application of deep learning model in the diagnosis of ECG signals to accurately identify two common cardiovascular diseases such as arrhythmia and Myocardial Infarction(MI).The main contents can be divided into the following points:Single-lead ECG signals were analyzed using a 1D convolutional neural network(CNN)to diagnose common arrhythmias.Traditional machine learning based solutions often require manual extraction and selection of features,while 1D CNN is the representative of deep learning.It can automatically capture key information in ECG signals through training without excessive manual intervention.At the same time,the CNN structure adopted in this paper is sufficiently lightweight,and it is implemented in combination with other algorithms on the System on a Chip(SoC),which can analyze and diagnose the ECG signal data stream in real time.Therefore,in addition to the PC platform,the algorithm is also suitable for wearable or portable devices with limited computing resources,and the application scenarios are more extensive.Moreover,based on the 4-lead ECG signal,a multi-lead CNN(Multilead-CNN,ML-CNN)was proposed to diagnose the highest-prevalence Generalized Anterior MI(GAMI).Multi-lead ECG signals have two characteristics: difference and integrity.Different leads have significant differences in the collected feature information because they are located at different distances and angles of the heart.However,all leads as a whole describe the current physiological characteristics of the same heart status.The ML-CNN is a modified model based on the characteristics of multi-lead ECGs,consisting of a specially designed sub-2D convolution and lead Asymmetric Pooling(LAP)to process a 2-dimensional matrix of multi-lead ECG signals stacked.Among them,the sub-2D convolution shares a 1D convolution kernel on different leads,which utilizes the integrity of the multi-lead ECG signal;while the LAP uses different pooling factors on different leads,emphasizing the diverse multi-lead characteristics.In the experiment,the ML-CNN is more suitable for multi-lead ECG signals.Diagnostic accuracy of ML-CNN for GAMI reached 96.00%,and the sensitivity and specificity were 95.40% and 97.37%,respectively.Furthermore,based on the 12-lead ECG signal,a Multiple Feature Branch-CNN(MFB-CNN)and its improved version are proposed to complete the diagnosis of a wider type of MI.The independent feature branch consists of a 1D convolutional layer and a pooled layer.It is in one-to-one correspondence with the lead,and the characteristics of each lead are extracted separately.Finally,the global fully connected layer is combined to determine the disease and reflect the integrity of the multi-lead..In its improved version of the MFB-Convolution Bidirectional Recurrent Neural Network(MFB-CBRNN),Long Term Memory(LSTM)is introduced to summarize the multi-lead information and Lead Random Mask(LRM)avoids overfitting and implements implicit integration,improving generalization capabilities.Both of them have achieved good results in class-based,patient-specific,and subject-based experiments based on categories,with accuracy rates above 93%.The work of this paper covers automatic diagnosis algorithms from single-lead,4-lead to all 12-lead ECG signals,and innovatively proposes improved CNN for multi-lead characteristics.In the case of fewer leads,the SoC deployment and real-time operation monitoring of the model is realized,and the application range is broadened.After effective integration and further improvement,the algorithm can play an active role in smart medical and auxiliary diagnosis in the future.