Unobtrusive Blood Pressure Estimation using Personalized Autoregressive Models

Cuffless and continuous blood pressure (BP) measurement using wearable devices is of great clinical value and health monitoring importance. Pulse arrival time (PAT) based technique was considered as one of the most promising methods for this purpose. Considering the dynamic and nonlinear relationship between BP, PAT and other cardiovascular variables, this paper proposes for the first time to use nonlinear autoregressive models with extra inputs (ARX) for BP estimation. The models were first trained by the baseline data of all 25 subjects to determine the model structure and then trained by individual data to obtain the personalized model parameters. To assess the effects of the dynamic and nonlinear factors, the data during water drinking and the first 5 minutes of recovery after drinking were used to validate the four models: linear regression, linear ARX, nonlinear regression and nonlinear ARX. The reference BP, which were measured by Finometer, were increased by 36.7±10.5 mmHg for SBP and 28.4 ±7.7 mmHg for DBP. This BP changes were best modelled by the nonlinear ARX, with Mean ± SD differences of 5.6 ± 8.8 mmHg for SBP and 3.8 ±5.8 mmHg for DBP. The study also showed that nonlinear factor significantly reduced the root mean square error (RSME) by about 50%, i.e., from 20.4 to 10.7 mmHg for SBP and 13.3 to 7.3 mmHg for DBP during drinking. While the effects of dynamic factors were not as significant as nonlinear factors, especially after introducing nonlinear factors.