Development of a blood pressure estimation hybrid deep learning system for wearable devices based on photoplethysmography

In this work, we developed a PPG-based blood pressure estimation hybrid deep learning model built into wearable devices and used by hypertension patients to monitor blood pressure in real-time in their daily lives. The model is a deep-learning model that combines data preprocessing, Autoencoder deep learning model for feature extraction, and RAN regression model developed by this research team. We conducted experiments to compare the blood pressure prediction performance of the proposed model with other deep learning models and find out how the objective blood pressure prediction performance is. We conducted experiments on an open dataset with the vital signs of 32 subjects. After models trained on 24 subjects' data and are tested on eight other people's data, we could see that using deep-learning regression models combined with an Autoencoder (hybrid deep-learning) performs better than using a deep learning model alone, and RAN accurately predicts blood pressure than the comparable deep-learning models. The study found that the average error for actual and predicted blood pressure in the proposed hybrid deep-learning models was 4.67 mmHg, and the standard deviation of error was 6.37 mmHg. It satisfies the accuracy criteria presented by the Korean National Institute of Food and Drug Safety Evaluation. © 2021 Korean Institute of Electrical Engineers. All rights reserved.