Research On Cardiovascular Disease Diagnosis Algorithms Based On Physiological Signals And Deep Learning

Cardiovascular disease is a class of high-risk diseases.It has now become a common and high-incidence disease that seriously threatens human life and health safety,and should attract wide attention from society.Electrocardiogram(ECG)signal is the most commonly used physiological signal in cardiovascular disease monitoring and diagnosis.Using computer technologies to assist in processing and analyzing ECG signals is significant for the monitoring and diagnosis of cardiovascular diseases,which is one of an important research and application directions at present.This thesis mainly uses deep learning methods to conduct research on cardiovascular disease monitoring algorithms based on ECG signals.For different current tasks and application scenarios,the main research and work contents of this thesis are as follows:(1)In this thesis,a multi-level dual-channel feature fusion ECG classification model is proposed for the task of single-lead ECG abnormality monitoring and diagnosis.To address the problems of poor interpretability and unsatisfactory recognition of single models in current research,the model extracts shallow features in the statistical,time and frequency domains of the ECG signal based on heart rate variability analysis,and the convolutional and recurrent neural networks were designed to extract deep spatial and temporal features of ECG signals in parallel in two channels respectively.The multi-level dual-channel feature fusion method achieves more effective utilization and complementary enhancement of ECG signal feature information,which enhances the classification accuracy and better classification result was achieved in experiments based on the Cin C 2017 single-lead ECG dataset.(2)In this thesis,a multi-lead ECG multi-label classification model based on multibranch multi-scale residual network is proposed,named MBMSML-ECGRes Net,for the task of multi-lead ECG multi-label abnormality classification.Based on the characteristics of multi-lead ECG signals,the model designs a multi-branch parallel network,combining 2D and 1D convolutions.Multi-scale convolutions are performed in parallel for multi-scale feature extraction and fusion in different branches.By designing improved residual units combined with channel spatial attention mechanism,the network structure is optimized and important feature representation is strengthened,achieving full mining and integration of feature information.Experiments on the CPSC 2018 multi-lead ECG dataset showed that the proposed model has better classification results.(3)In this thesis,a model for ECG signal data generation based on self-attention generative adversarial networks is proposed for the task of ECG signal generation and data enhancement.To improve the quality of generated data to effectively address data imbalances,a generator model based on Bi-LSTM combined with CNN is designed to integrate the temporal and spatial features of ECG data to achieve better generation results,and a discriminator model based on CNN combined with minibatch discrimination is also designed to suppress mode collapse problems and ensure the diversity of generated ECG signal data.Furthermore,a self-attention-mechanism-based ECG sequence feature enhancement module is designed to strengthen feature representation and better guide the generation of ECG signal data.The effectiveness of the proposed method was verified through comparative experiments on the Cin C 2017 single-lead ECG dataset.