Comparison of Three Novel Machine Learning Models for Photoplethysmography-based Cuffless Blood Pressure Prediction

Photoplethysmography-based non-invasive blood pressure (NIBP) monitoring is critical for cardiovascular disease prevention and health management, as it enables continuous and real-time assessment without causing discomfort to the patient. We collected 71 time-domain features, 16 frequency-domain features, and 29 demographic features from 809 subjects. Each subject underwent three 120-second Photoplethysmography (PPG) acquisition sessions. After each session, systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured using the Omron HEM-7312 sphygmomanometer. Subsequently, these PPG segments were used to predict the blood pressure (BP) by the traditional SVM model and three novel machine learning models: particle swarm optimization-optimized SVM (PSO-SVM), radial basis function neural network (RBFNN), and random forest (RF). The performance of each model was evaluated. The experimental results demonstrated that the RF model achieved the optimal performance in predicting both SBP and DBP, with RMSE values of 1.165 mmHg and 0.928 mmHg, respectively. Additionally, among the three novel machine learning models, RF exhibited higher computational efficiency. This study provides a theoretical basis for selecting appropriate algorithms in blood pressure monitoring.