A deep learning method for continuous noninvasive blood pressure monitoring using photoplethysmography

Abstract

Objective. The aim of this study is to investigate continuous blood pressure waveform estimation from a plethysmography (PPG) signal, thus providing more human cardiovascular status information than traditional cuff-based methods. Approach. The proposed method utilizes the feature extraction ability of a convolution neural network to estimate blood pressure (BP) from PPG signals without the need for waveform analysis and signal feature extraction. Main results. The network achieved mean absolute errors and standard deviations of 2.55 ± 3.92 mmHg for systolic BP (SBP), 1.66 ± 2.76 mmHg for diastolic BP (DBP), and 2.52 ± 3.02 mmHg for overall pressure waveform. The results meet the best levels of the protocols of the British Hypertension Society (BHS) and the Association for the Advancement of Medical Instrumentation (AAMI). Significance. The proposed method shows promise for noninvasive continuous BP monitoring in hospital wards and daily life, which can assist in clinical diagnosis, disease treatment, and rehabilitation.