Research And Application Of ECG Aided Diagnosis Based On Lightweight Neural Network

Cardiovascular disease(CVD)is one of the most important diseases that threaten human life and health,and some cardiovascular diseases such as atrial fibrillation(AF)and myocardial infarction(MI)are prone to increase the risk of cerebral infarction if not treated timely,It is necessary and urgent to prevent and treat such cardiovascular diseases as soon as possible.The Electrocardiogram(ECG)real-time monitoring and diagnosis system can analyze ECG signals in a long-term and real-time manner,and make timely early warnings for the onset of cardiovascular diseases,so as to carry out effective treatment and intervention.The over-parameterization and redundant characteristics of neural networks lead to expensive computational costs.The realization of a lightweight ECG-assisted diagnosis model with high accuracy is the main challenge faced by current ECG real-time monitoring and diagnosis systems.This thesis starts from the demand of real-time ECG monitoring and diagnostic system,and takes AF and MI,two extremely high-risk cardiovascular disease detection as the specific application scenarios,with the research goal of achieving lightweight and high-accuracy ECG-assisted diagnostic models.The research focuses on two aspects:lightweight high-performance detection model and energy-efficient ECG realtime monitoring and diagnosis system,respectively at the algorithm and system level,to achieve real-time auxiliary analysis and diagnosis of very high-risk cardiovascular diseases and make timely warning.The main research work includes:(1)The application of convolution neural network(CNN)in real-time detection of AF has the problems of large amount of parameters and high computational complexity.To address this problem,a lightweight automatic detection model of AF based on depthwise separable convolution is constructed.By combining pooling and depthwise separable convolution,the depthwise separable max-pooling convolution module is designed to reduce the number of convolution operations,At the same time,aiming at the problem of signal feature loss in pooling operation,a lightweight convolutional block attention module(CBAM)is introduced to optimize the loss function,and screen important feature maps.The experimental results show that the model parameter size is 0.054MB,FLOPs is 14.13M,and achieved 96.74%on the MITBIH Atrial Fibrillation Database(AFDB).It can effectively improve the detection efficiency while ensuring high classification accuracy and meet the real-time detection requirements of AF.(2)To address the problem of extracting MI feature information contained in multiple leads for real-time MI detection tasks,the dense connections of the densely connected convolutional neural network(DenseNet)increase the number of convolutional operations while enhancing the reuse of ECG signal features between leads.Firstly,DenseNet is built to improve the reuse of feature information,and then the network pruning technique is used to evaluate the importance of filters in each weighted layer of DenseNet by using parametric numbers,and the pruned filters are continuously updated during the training phase.The experimental results showed that the accuracy,sensitivity and specificity of the model was 97.76%,96.65%and 97.78%on the Physikalisch Technisch-Bundesanstalt(PTB)database.Compared with the pretrained model,the number of parameters was reduced by 32%,meeting the requirements for real-time MI detection.(3)Establishing a remote ECG real-time monitoring and diagnosis system.According to the different needs of users for the system,the ECG real-time monitoring and diagnosis system is designed and implemented,and the functions of ECG data realtime transmission and storage,ECG data analysis and processing,user information management and ECG data real-time display are completed.The system was tested for functionality and performance,and the test results showed that the system can achieve real-time analysis and diagnosis of risky cardiovascular diseases.