Comparision of BP Estimation Model

Blood pressure (BP) is a vital parameter in medical treatment and diagnosis, and as a non-invasive method to measure BP, some deep learning models have been proposed to estimate BP from photoplethysmograms (PPGs). However, the datasets and the method of dividing them into training and testing subsets are not uniform, making it difficult to compare them fairly. In this study, we compared the performance of nine deep learning models for estimating systolic and diastolic BP using PPGs. We used PulseDB, which has "calibration-based subset'' and "calibration-free subset'' as test subsets. The calibration-based subset has the same subject's data in the training subset, and the calibration-free subset does not have the same subject's data in the training subset. The results showed that ST-ResNet performed the best, and it is important to evaluate the models using both calibration-based and calibration-free subsets, and to prevent the overfitting by using the weight decay.