Deep Learning Model for Blood Pressure Estimation from PPG Signal

Minseong Kim, Hyeonjeong Lee, Kwang-Yong Kim, Kyu-Hyung Kim

Abstract: In this study, we propose a deep learning-based framework to estimate blood pressure using photoplethysmogram (PPG) signals. We also propose a calibration method that applies the initial blood pressure information to the estimated results. To evaluate our approach, we used the PPG and blood pressure signals of 4200 patients sampled from the MIMIC-III Waveform Database. The resulting mean absolute error and standard deviation were 4.876mmHg and 5.257mmHg, respectively. Compared to the case of not calibrating using initial blood pressure information, we achieved the performance improvement of mean absolute error of 1.899mmHg and standard deviation of 2.933mmHg.