Cuffless blood pressure estimation from PPG signals and its derivatives using deep learning models

Abstract

Blood pressure (BP) is a direct indicator for hypertension, therefore, continuous and non-invasive BP monitoring is essential for reducing future health complications. Most non-invasive blood pressure monitors use the oscillometric technique, which can be cumbersome and impractical. To address this problem, we explore several features extracted from the Photoplethysmogram (PPG) waveform and its first and second derivatives and employ deep learning recurrent models for non-invasive cuffless estimation for systolic and diastolic BP. In this research, three techniques have been considered including statistical and machine learning techniques for reducing the collinearity and redundancy in the input feature vector. The estimation models consist of a one bidirectional recurrent layer, followed by a series of stacked conventional recurrent layers and an attention layer. All models were evaluated on 942 subjects collected from the MIMIC II dataset. The best performing model (consists of one bidirectional layer followed by several Long Short-Term Memory layers and an attention layer) achieved a mean absolute error, and standard deviation of 4.51 ± 7.81 mmHg for systolic BP (SBP), and 2.6 ± 4.41 mmHg for diastolic BP (DBP). The results show that the deep learning model trained on features extracted from one PPG sensor yield good performance for continuous and cuffless BP monitoring. Additionally, the results fulfil the international standard for cuffless BP estimation.