Tuesday, November 15, 2016

Systolic blood pressure estimation using PPG and ECG during physical exercise.

Systolic blood pressure estimation using PPG and ECG during physical exercise.

Physiol Meas. 2016 Nov 14;37(12):2154-2169

Authors: Sun S, Bezemer R, Long X, Muehlsteff J, Aarts RM

In this work, a model to estimate systolic blood pressure (SBP) using photoplethysmography (PPG) and electrocardiography (ECG) is proposed. Data from 19 subjects doing a 40 min exercise was analyzed. Reference SBP was measured at the finger based on the volume-clamp principle. PPG signals were measured at the finger and forehead. After an initialization process for each subject at rest, the model estimated SBP every 30 s for the whole period of exercise. In order to build this model, 18 features were extracted from PPG signals by means of its waveform, first derivative, second derivative, and frequency spectrum. In addition, pulse arrival time (PAT) was derived as a feature from the combination of PPG and ECG. After evaluating four regression models, we chose multiple linear regression (MLR) to combine all derived features to estimate SBP. The contribution of each feature was quantified using its normalized weight in the MLR. To evaluate the performance of the model, we used a leave-one-subject-out cross validation. With the aim of exploring the potential of the model, we investigated the influences of the inclusion of PAT, regression models, measurement sites (finger and forehead), and posture change (lying, sitting, and standing). The results show that the inclusion of PAT reduced the standard deviation (SD) of the difference from 14.07 to 13.52 mmHg. There was no significant difference in the estimation performance between the model using finger- and forehead-derived PPG signals. Separate models are necessary for different postures. The optimized model using finger-derived PPG signals during physical exercise had a performance with a mean difference of 0.43 mmHg, an SD of difference of 13.52 mmHg, and median correlation coefficients of 0.86. Furthermore, we identified two groups of features that contributed more to SBP estimation compared to other features. One group consists of our proposed features depicting beat morphology. The other comprises existing features depicting the dicrotic no! tch. The present work demonstrates promising results of the SBP estimation model during physical exercise.


Read More