Research On Deep Learning-based Assisted Diagnosis Method For Cardiovascular Diseases

Cardiovascular disease is one of the leading causes of mortality worldwide.Early diagnosis and treatment are crucial for improving patient survival rates.However,traditional methods for diagnosing cardiovascular disease suffer from high misdiagnosis rates and low detection accuracy,making early and accurate diagnosis challenging.This thesis explores the possibility of automated assisted diagnosis of cardiovascular disease by employing deep learning techniques and combining electrocardiogram signals with multidimensional structured cardiovascular disease data.Through this approach,more precise early prevention and treatment guidelines can be provided for patients.The main research tasks accomplished are as follows:Firstly,a heart signal automatic diagnostic classification model,MECNN-GRU,based on convolutional neural network and gated recurrent unit,is proposed for heart signal data.Noise affects accuracy in the classification process of heart signals,so the signal is first scaled by Daubechies4(db4)wavelet function,and appropriate threshold functions are used to filter the wavelet coefficients of each scale to remove noise components,and the signal is reconstructed to obtain the denoised heart signal.Then,the preprocessed data is input to the MECNN-GRU model for heart beat classification.In this model,a multi-head convolution module is used to fully extract the spatial features of the heart signal from different scales using different sizes of convolution kernels.Then,the GRU module is used to extract the temporal features of the heart signal.In addition,an autoencoder is constructed to extract latent feature information from the header annotations of the MIT-BIH database.The experimental results show that the overall accuracy of the constructed MECNN-GRU model for heart beat classification reached99.02%.Secondly,a diagnostic classification model,CVDCNN,based on residual convolutional neural networks,was proposed for multi-dimensional structured cardiovascular disease data.Firstly,missing and abnormal values in the multidimensional cardiovascular disease data were processed,and the interaction between features was identified through data visualization.Then,the preprocessed data was input into the CVDCNN model for classification diagnosis.In the CVDCNN model,the feature dimension was increased through a fully connected layer,and more non-linear transformations were introduced to improve feature representation ability.The data was then reshaped into a multi-channel form to allow the model to better learn the correlations between multi-dimensional features of cardiovascular disease data.Additionally,R-Drop was used in the model to alleviate the inconsistency between model training and inference caused by Dropout,and the SHAP framework was used for explanatory analysis to enhance the model’s interpretability.The experimental results showed that the constructed CVDCNN model achieved an AUC value of 80.05% for multi-dimensional structured cardiovascular disease data classification,which was better than the listed comparative methods.The key factors affecting cardiovascular disease were identified using SHAP,demonstrating the effectiveness of the proposed method in this thesis.