Comparison of seven shallow and deep regressors in continuous blood pressure and heart rate estimation using single-channel photoplethysmograms under three evaluation cases

Abstract

Our objective is to compare seven blood pressure (BP) and heart rate (HR) estimation models based on shallow and deep regressors using single-channel photoplethysmograms (PPGs) under three evaluation cases. Non-invasive, cuffless, continuous, and simultaneous estimation of systolic BP, diastolic BP, and HR using only PPGs is valuable for routinely monitoring the health status using simple device configuration. However, the performances of regressors cannot be directly compared based on the metrics presented in their respective papers because they use different datasets and evaluation procedures. Therefore, we used a common dataset (containing 1811 patients) and prepared three evaluation cases in which the training and testing data included, partly included, and did not include data from the same subjects. Shallow models using hand-crafted features performed reasonably well, and when the training data contained data similar to the testing data, deep models using raw signals as input achieved accurate BP and HR estimation. The spectral–temporal residual network showed the highest performance, with modified PP-Net as the runner-up; this study revealed that they are strong candidates for non-invasive, cuffless, continuous, and simultaneous BP and HR estimation models using single-channel PPGs in clinical and daily-use scenarios. Furthermore, results showed that the estimation performances rely on the level of inclusion of same subjects’ data in the training and testing data. To address the discrepancy between pre-measured data domain and new user’s data domain, we will investigate effective transfer learning techniques in the future work. The training and testing datasets used in this study are made public (https://drive.google.com/drive/folders/14nSeg0V-metCIs3LHPBH9tolxiyCN8eA?usp=share_link).