Blood Pressure Estimation from PPG Signals Using Deep Residual Network with Transfer Learning

In this study, we present an approach that enables transfer learning-based estimation of systolic and diastolic blood pressure (SBP and DBP) using photoplethysmography (PPG) signal. In the development of BP estimation models, we use the MIMIC II database containing PPG and arterial BP signals. The proposed approach utilizes algorithms based on recurrent neural network architecture with ResNet50, VGG16 and MobileNetV2 deep features. The results show that a high estimation performance is achieved with the selected ResNet50 deep features and bi-directional gated recurrent unit algorithm. Our proposed approach has a mean absolute error and standard deviation (MAE +/- SD) of 5.66 +/- 8.82 and 2.82 +/- 5.60 mmHg for estimating SBP and DBP, respectively. The proposed estimation model satisfies both the BHS and AAMI standards for SBP and DBP. The results demonstrate the effectiveness of the proposed approach to estimate BP, especially in patients at risk for cardiovascular disease and hypertension.