Research On Non-invasive Continuous Blood Pressure Estimation Approach Based On Machine Learning

Considering the existing issues of traditional blood pressure measurement methods and noninvasive continuous blood pressure measurement techniques,this study established the systolic blood pressure(SBP)and diastolic blood pressure(DBP)estimation models based on machine learning using pulse transit time(PTT)and characteristics of pulse waveform.In the process of model construction,the mean impact value(MIV)method was introduced to investigate the impact of each feature on the models.And the genetic algorithm(GA)was introduced to implement parameter optimization.The experiment results showed that the proposed models had higher model accuracies than the traditional methods,and can effectively describe the nonlinear relationship between the features and blood pressure.The estimation errors met the requirements of AAMI and BHS criteria.Our study is helpful to promote the practical application of methods for noninvasive continuous blood pressure estimation models.The main research work and results of this paper include:(1)Laboratory blood pressure experiments were designed to collect blood pressure data from healthy young people and to supplement waveform data including electrocardiogram(ECG),photoplethysmography(PPG)and arterial blood pressure from the MIMIC III waveform database.(2)The denoising algorithm of ECG signal is designed based on wavelet threshold denoising method.The denoising algorithm of PPG signal is designed based on the combination of wavelet threshold denoising method and cubic spline interpolation method.Biorthogonal quadratic B-spline wavelet is introduced in recognizing the R wave of ECG,and the sliding window method,the differential method and the curvature method is introduced in identifying the pulse wave characteristics.(3)A blood pressure estimation model based on the SVR method in machine learning is proposed.After normalizing the collected features,the MIV method is used to explore the influence of features on the model and to screen out the less influential features.In the process of training the model,GA optimization is introduced to optimize the parameters,which effectively avoids the state of the model falling into the state of over-learning or under-learning.(4)The model’s predictive ability is verified by test samples,and Bland-Altman is used to analyze the consistency between predictive and actual values.At the same time,compared with the model based on multivariate linear regression and PTT-SVR methods,the proposed model method has high accuracy and strong predictive ability,and it is feasible in practical daily non-invasive continuous blood pressure measurement.