Research On Convolutional Neural Networks Applied To Arrhythmia Classification And Continuous Blood Pressure Monitoring

Globally,cardiovascular disease(CVD)is now the leading cause of death.There are many different types of cardiovascular diseases,and arrhythmia and hypertension are two of their key symptoms.Palpitations,tightness in the chest,syncope,and rapid cardiac death can result from arrhythmia.Blood pressure that is continuously higher than normal against vessel walls is a condition known as hypertension.Increased cardiac load,an increased risk of atherosclerosis,the development of various diseases,and more are all effects of hypertension.Currently,clinicians mostly use ECG records to detect arrhythmia condition,and wrist blood pressure readings are the standard approach for blood pressure monitoring.The aforementioned procedure is difficult and time-consuming,has poor measurement time control,and is susceptible to human error.In order to classify arrhythmias and continuously non-invasive blood pressure monitoring,the deep learning method was applied in this thesis.The specific research included in this thesis is:(1)Study of arrhythmia classification automatically using ECG signals.The electrocardiogram(ECG)represents the electrical activity of the heart’s circuit and documents the process of creating,spreading,and recovering cardiac excitement.Arrhythmia can be diagnosed objectively based on a rhythm abnormality and a change in the waveform of the ECG.Traditional methods usually ignore the temporal characteristics of ECG signals as time series and lack of enough training samples of different cardiac rhythm types in ECG signals.In this thesis,a data augmentation method combining Gramian angular summation field and generative adversarial network is proposed to maintain the class balance,and then a deep residual convolutional neural network with attention mechanism is constructed as a classification network.The algorithm was verified on the MIT-BIH arrhythmia dataset,and the results show that the proposed method can successfully achieve the automatic classification of arrhythmia type.(2)Study of continuous non-invasive blood pressure monitoring using PPG signals.The photoplethysmography(PPG)signal is caused by periodic fluctuations in blood flow,blood pressure,and vessel walls propagating through the arterial system as the heart diastole and contract intermittently.Using PPG signal is a promising method to blood pressure due to the association between PPG signal and blood.In this thesis,a continuous non-invasive blood pressure monitoring method based on multi-scale residual U-shaped network(SE-MSResUNet)integrated with squeeze and excitation module is proposed,and blood pressure is estimated indirectly by monitoring arterial blood pressure waveform.When the SE-Inception module is introduced into the coding path to extract multi-scale features,the model focuses on the relationship between feature channels and automatically learns the importance of feature channels.In addition,the residual path is selected to replace the original skip connection in the connection mode of encoding path and decoding path,so as to solve the problem of feature incompatibility caused by the direct fusion of low-level features and high-level features.The algorithm was verified on MIMIC dataset,and the results show that the proposed method can successfully achieve continuous non-invasive blood pressure monitoring.The above work in this thesis provides new insights for cardiovascular status assessment from the two aspects of arrhythmia classification and continuous blood pressure monitoring,and provides a new solution for the realization of early prevention,diagnosis and treatment of cardiovascular diseases,making it possible for individuals to wear and portable cardiovascular disease health monitoring devices.