Study Of Blood Pressure Prediction Model Based On Photoplethysmography

Globally,hypertension is a major contributor to cardiovascular disease(CVD)and death.Because hypertension is often asymptomatic in its early stages,many people do not receive timely treatment.It is simple to obtain and suitable for long-term monitoring to use photoplethysmography(PPG).Therefore,early diagnosis and management of hypertension using PPG signals is important for the prevention and treatment of CVD.In order to create a real-time artificial intelligence-based blood pressure prediction system,we design hypertension diagnosis algorithms based on machine learning and deep learning technologies for various processing of gathered PPG data.The main research includes:(1)Analysis was done on how the pulse wave and blood pressure are formed as well as their affecting elements.And the PPG data required for the study were collected through two intensive care medical information databases,sliced into 5 s fixed-length signal segments,and the synchronised blood pressure data were used to mark the PPG signal segments.(2)Firstly,for the small dataset,the velocity pulse wave and acceleration pulse wave are obtained by differential processing of PPG signal,and the features are extracted and filtered using Tsfresh,and the effective features are used as the input of Light GBM classifier tuned by Optuna method for machine learning hypertension risk stratification experiments;Additionally,the filtered PPG signal for the full dataset is For more precise deep learning blood pressure numerical prediction experiments,the time-frequency chromatogram derived by continuous wavelet transform is further employed as the input of the combined convolutional neural network and bidirectional long and short-term memory network model;Ultimately,by implementing the learnt model on the hardware platform,a deep learning-based blood pressure prediction system is realized.(3)For the experimental results of machine learning hypertension risk stratification,the F1 scores were higher than 90%,and the diagnostic performance was better than other machine learning classifiers,and the combination of PPG signal and its derivatives could improve the diagnostic performance compared with the hypertension stratification experiments using only a single signal.For the deep learning blood pressure numerical prediction experiments,the mean absolute error ± standard deviation for systolic and diastolic blood pressure were 3.80 ± 5.02 mm Hg and 1.65 ± 2.70 mm Hg,respectively,which met the standards of the American Medical Devices Association and the British Hypertension Society,and the time-frequency conversion could retain the valid information of PPG to a greater extent compared with other signal conversion methods.This paper investigates hypertension diagnosis using PPG signals by machine learning method and deep learning method,respectively,and the results show that the two methods show high accuracy in hypertension risk stratification and blood pressure prediction,providing a non-invasive,fast and stable method for the early detection of hypertension with a wide range of potential applications for wearable cuffless blood pressure measurement.