Personalized Modeling of Blood Pressure with Photoplethysmography: an Error-Feedback Incremental Support Vector Regression Model

• Published in: IEEE internet of things journal Vol. 11; no. 1; p. 1
• Main Authors: Wang, Dingliang, Yang, Xuezhi, Wu, Jun, Wang, Wenjin
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptation models; Biomedical monitoring; Blood pressure; blood pressure (BP); Customization; Data models; Edge computing; EFISVR; Error feedback; Estimation; Feature extraction; incremental learning; Internet of Things; Lightweight; Monitoring; personalization; photoplethysmography (PPG); Regression models; Support vector machines; support vector regression (SVR)
• ISSN: 2327-4662; 2327-4662
• DOI: 10.1109/JIOT.2023.3290557

Most of the existing photoplethysmography (PPG) based blood pressure (BP) estimation methods aim at training a general BP model applicable to all individuals which neglected the vasculature and anatomical differences among individuals as well as the slow and subtle cardiovascular changes over time, thus were hard to achieve high accuracy. This study aims at addressing this problem by constructing personalized BP models from PPG signals. First, the PPG features that can well reflect the changes of individual BP with physiological state were extracted. Afterwards, an Error Feedback Incremental Support Vector Regression (EFISVR) model was designed to achieve high-accuracy BP measurement of a subject, which can quickly be adapted to new samples without retraining the whole model. Results show that the constructed model can accurately predict the BP values of a subject for at least 3 months. The mean absolute error (MAE) of BP estimation were 3.11 mmHg for systolic blood pressure (SBP) and 2.47 mmHg for diastolic blood pressure (DBP). The proposed EFISVR model is lightweight which can be integrated into wearables and other edge devices, as a part of Internet of Things (IoT) applications. The advantages of lightweight, few-shot learning and high-precision make the model suitable for applications in real-life scenarios.