Continuous Blood Pressure Estimation from PPG Signal

Abstract

Given the importance of blood pressure (BP) as a direct indicator of hypertension, regular monitoring is encouraged for healthy people and mandatory for patients at risk from cardiovascular diseases. We propose a system in which photoplethysmogram (PPG) is used to continuously estimate BP. A PPG sensor can be easily embedded in a modern wearable device, which can be used in such an approach. A set of features describing the PPG signal on a per-cycle basis is computed to be used in regression models. The predictive performance of the models is improved by rst using the RReliefF algorithm to select a subset of relevant features. Afterwards, personalization of the models is considered to further improve the performance. The approach is validated using two distinct datasets, one from a hospital environment and the other collected during every-day activities. Using the MIMIC hospital dataset, the best achieved mean absolute errors (MAE) in a leave-one-subject-out (LOSO) experiment were 4.47 +- 5.85 mmHg for systolic and 2.02 +- 2.94 mmHg for diastolic BP, at maximum personalization. For everyday-life dataset, the lowest errors in the same LOSO experiment were 8.57 +- 7.93 mmHg for systolic and 4.42 +- 3.61 mmHg for diastolic BP, again using maximum personalization.