| With the development of society,people’s living standard has gradually improved.Bad habits such as staying up late and uncontrolled diet have increased the incidence of hypertension year by year.Under the dual pressure of aging population and metabolic risk factors,cardiovascular diseases caused by hypertension have become a major burden in China.To reduce the prevalence of cardiovascular diseases,monitoring and prevention of hypertension has become important,and continuous blood pressure monitoring has become the focus of research in recent years.Continuous blood pressure monitoring can be achieved using the pulse wave characteristic parameter method,which fits the current demand for blood pressure measurement.This blood pressure measurement method is based on photoplethysmographic(PPG),which is commonly used in watches,bracelets,and other wearable devices,and can be adapted to the blood pressure measurement needs of different scenarios,and has the advantages of simple operation and wide applicability to the population,effectively increasing the awareness of hypertension.However,blood pressure is complex and diverse,and blood pressure rhythms can be classified into arytenoid,non-arytenoid,super-arytenoid and anti-arytenoid according to different diurnal patterns,and the rhythmicity of blood pressure is also affected by seasons,and these characteristics often lead to insufficient measurement accuracy of the trained models.In this paper,we propose three time-dependent characteristic parameters to optimize the measurement accuracy of the model,and integrate the classification algorithm to build a blood pressure estimation model and investigate the effect of the classification algorithm on the accuracy of blood pressure measurement.The main contents of this paper include the following aspects.(1)135 PPG time-domain features and 399 derived features were successfully extracted from the original PPG signal and its second-order derivative signal.Specifically,the PPG signal is first preprocessed using Butterworth filter to locate the crest and trough points,extract the effective bands in the PPG signal,locate the maximum slope point along the rising edge,the repetition wave point and the four extreme points of the secondorder derivative signal by using the interval method,and then calculate the required timedomain features and derived features based on the extracted feature points.(2)A blood pressure estimation model conforming to the Association for the Advancement of Medical Instrumentation(AAMI)standard was developed based on the PPG features selected in this paper.The PPG features selected by the filtering method were used to build a gradient boosting decision tree and random forest blood pressure estimation model,and the validity was verified,and then the features were selected again using Recursive Feature Elimination(RFE),Genetic Algorithm(GA)and embedding method.The results showed that the model based on the features selected by RFE could obtain high accuracy and the measurement results could meet the AAMI criteria,and the mean error and standard deviation of the measured Systolic Blood Pressure(SBP)could reach-0.114 mm Hg and 7.950 mm Hg,and the mean error of the measured Diastolic The mean error and standard deviation of the measured diastolic blood pressure(DBP)can reach 0.032 mm Hg and 4.937 mm Hg.(3)Three time-dependent characteristic parameters were proposed to optimize the measurement accuracy of the model,and a blood pressure estimation model conforming to the AAMI standard and the British Hypertension Society(BHS)standard was established.Based on the PPG characteristics selected by RFE,the blood pressure estimation model was built by gradually fusing the measurement time(month,hour),historical PPG characteristics,and historical blood pressure data.The results showed that using blood pressure measurement time as a feature could effectively improve DBP measurement accuracy,and the mean absolute error and standard deviation of measured DBP decreased by up to 9.01% and 7.75%.After incorporating historical PPG features as well as historical blood pressure data on the basis of temporal features,the model significantly improved the measurement accuracy of both SBP and DBP,and the mean absolute error and standard deviation of measured SBP and DBP decreased by up to 11.66%and 8.90%,and the standard deviation decreased by up to 10.90% and 10.23%.After the hyperparameters were Bayesianoptimization(BO),the SBP measurements reached the B-class standard of BHS,and the DBP measurements reached the A-class standard of BHS.(4)A blood pressure estimation model incorporating classification algorithms was constructed to investigate the effect of classification algorithms on blood pressure measurement accuracy.Support Vector Machine(SVM)and Extreme Gradient Boosting(XGboost)classification models were built based on the features selected by the filtering method,and the features were selected using RFE and GA,and the model classification performance was further improved using Bayesian optimization algorithm.BO-GAXGboost model achieves more than 82.75% for hypertension data,more than 90.31% for other blood pressure data,and more than 89.91% accuracy.The blood pressure estimation model of fused BO-GA-XGboost classification algorithm was constructed and compared with the measurement results of fused ideal classification algorithm.The results showed that the fusion classification algorithm could improve the measurement accuracy of the model,but because there was still a gap between the accuracy and completeness of the BO-GA-XGboost model and the ideal classification model,the mean absolute error and standard deviation of the measurements appeared to increase,although the percentage of SBP measurement results with absolute errors not exceeding 5 mm Hg increased.This paper explores methods to improve the accuracy of model measurements at the feature level and algorithm level,providing a theoretical basis for achieving accurate continuous blood pressure monitoring in the future. |
PDF Full Text Request