Research On Continuous Non-invasive Blood Pressure Measurement Based On PPG And ECG Signals

Blood Pressure(BP)is not only an important indicator that must be measured in routine physical examinations,but also a key monitoring parameter of the cardiovascular system during cardiac surgery,drug testing,and intensive care.It can timely reflect the effect of cardiovascular interventional surgery and drug antihypertensive effect to provide scientific reference for medical staff.The correlation between Pulse Transit Time(PTT)and BP is the most common method used by predicting BP without a cuff.However,which is limited due to the following problem:(1)The morphological contours of Photoplethysmograph(PPG)and Electrocardiogram(ECG)signals will change due to the interaction of noise and other physiological systems,and extracting features from ECG and PPG signals and calculating PTT is very difficult;(2)The correlation between PTT and BP varies from person to person and needs to be calibrated regularly.Driven by medical big data,previous studies have shown that the application of deep learning in non-invasive continuous BP prediction is feasible and promising.With the excellent feature extraction capabilities of deep learning,it can provide a theoretical basis for portable blood pressure measuring devices to achieve high-precision and continuous blood pressure measurement of people.Therefore,continuous non-invasive measurement BP method was proposed in this study based on deep convolutional neural network,using multiple cycles PPG and ECG waveform signals without feature engineering.The main research contents of this paper are as follows:1)Study the quality assessment methods of ECG,PPG and BP signals.Accurate correspondence between input features and labels is the premise of deep learning modeling.An ECG signal quality assessment method based on signal kurtosis and R-peak matching is proposed.For PPG and ABP signals,a method to transform them into pictures is proposed for quality evaluation.A data set of PPG signal quality images was built,and twodimensional convolutional neural network was used to classify and evaluate it.The results show that the signal quality evaluation can effectively eliminate the invalid data and retain the richness of the original sample,which lays a foundation for the accurate modeling of blood pressure estimation.2)A continuous non-invasive blood pressure estimation method was studied based on the combination of recurrent neural network(RNN)and convolutional neural network(CNN).Traditional non-invasive continuous blood pressure measurement method based on machine learning is limited due to manual extraction of pulse wave feature lacking expression ability,poor generalization ability and periodic calibration.This study proposes a continuous BP measurement method based on deep learning,which uses multicycle PPG and ECG signals to reflect the inherent characteristics of different individuals.The convolutional neural network was used to extract the detail features from the original ECG and PPG signals,and the channel attention was adopted to allocate the adaptive weight of the extracted features to improve the prediction accuracy of the model.3)Comparison experiments with different super parameters were designed to get the model with the best performance.Finally,the prediction results of the best model under the calibration and uncalibration methods were compared with other studies.Before calibration,the prediction accuracy of systolic blood pressure(SBP)was 10.89±7.12 mm Hg,and that of diastolic blood pressure(DBP)was 5.79±3.60 mm Hg.After calibration,the prediction accuracy of SBP was 4.70±3.45 mm Hg,and the prediction accuracy of DBP was2.40±2.99 mm Hg,which proved that the model had a high accuracy in the non-invasive continuous measurement method of blood pressure.