Predicting Blood Pressure from Photoplethysmogram Data: A Machine Learning Approach

The prediction of BP from photoplethysmogram (PPG) data using machine learning has garnered significant interest. However, discrepancies in dataset size, subject numbers, and pre-processing methods across studies hinder fair performance comparisons and obscure the generalization capabilities of BP estimation models. To address this, the largest dataset available for benchmarking BP estimation models “PulseDB”, was introduced, and designed to meet standardised testing protocols. PulseDB is sourced from both the VitaIDB database and the MIMIC-III waveform database. It also provides subjects' identification and demographic information for enhanced model performance and generalizability assessment, also PPG signal characteristic points for minimal pre-processing in deep learning models. Using PulseDB, this study is the first to investigate the performance gap between calibration-based method & the calibration-free testing methods regarding the generalizing ability of blood pressure estimation models. The PulseDB database aims to serve as reliable, user friendly, and comprehensive data for evaluating BP estimation methods. This study presents a blood pressure measuring system using photoplethysmograph signals and machine learning (ML) models and identifies optimal regression models for independent estimation of the SBP and the DBP. This model yielded significant outcomes, achieving the RMSE of 8.7955 for SBP and 5.3977 for DBP, highlighting its superior performance in BP estimation.