Research On Auxiliary Diagnosis Of Cardiovascular Disease Based On LSTM

Cardiovascular Disease(CVD)accounts for 45.01% and 42.61% of deaths in rural and urban areas,respectively.Electrocardiogram(ECG)is the main way to diagnose cardiovascular diseases.However,Short-term ECG is difficult to detect sudden abnormalities,and long-term monitoring equipment is time-consuming and inefficient.Based on telemedicine,automatic analysis of ECG has theoretical significance and application value for solving the above problems.Using prior knowledge and deep learning to auxiliary diagnosis is the two main technical routes for ECG automatic analysis.Auxiliary diagnosis based on prior knowledge relies on ECG feature extraction.Individual differences in ECG lead to uncertainty in feature extraction,which reduces the accuracy of algorithm classification.The auxiliary diagnosis based on deep learning is data-oriented and does not depend on the feature extraction process,which provides a feasible technical approach for ECG big data analysis.Aiming at remote diagnosis of cardiovascular disease,based on large-scale ECG data sets,an improved long short-term memory network is used to establish a single-task and positive anomaly auxiliary diagnosis model to solve the problem of feedforward neural network which is difficult to learn ECG timing features.Furthermore,by mining disease correlation,a multi-task and multi-classification auxiliary diagnosis model is constructed to increase the types of disease recognition.Finally,a remote ECG monitoring system based on cloud platform is designed to realize remote auxiliary diagnosis function.The main research work includes the following:(1)Using the memory function of Long Short-Term Memory(LSTM)and the receptive field characteristics of Convolutional Neural Networks(CNN),the spatial and temporal characteristics of ECG are learned autonomously,and the CNN-LSTM hybrid model is constructed.First,the ECG is preprocessed using a notch filter and an adaptive threshold algorithm based on wavelet transform.Then,CNN is used to capture the spatial characteristics of each ECG lead and LSTM is used to mine the temporal correlation between ECG signal points.Finally,based on experimental and clinical databases,the model classification accuracy rates were 99.6% and 93.39%,respectively.(2)Using the spatial and temporal characteristics of the hybrid network of CNN and Bidirectional Long Short-Term Memory(Bi-LSTM),Combining the attention mechanism,selectively learning the input data and using Multi-Task Learning(MTL)to classify and identify multiple diseases at different levels.A CNN-BiLSTM hybrid model is constructed to solve the problem that single-task can easily neglect the correlation information between diseases,and the single classification can only identify one disease.First,the data set is divided by the hold-out method.Each ECG marks two labels as the two-dimensional matrix input,and then uses the CNN-BiLSTM network to mine the spatial and temporal characteristics of the ECG,and outputs the results according to the joint weights to achieve multiple disease classification.Finally,based on 170,000 clinical data analysis,the model classification accuracy rate reached 88.01%.(3)Designing a remote ECG monitoring system with a three-layer structure,which included data acquisition equipment,cloud platform and client software,it implemented data collection,data analysis and data display functions.The model is used to realize remote cardiovascular disease diagnosis.The test results showed that the system can realize remote cardiovascular disease diagnosis function and verify the feasibility of the algorithm.