Hybrid D1DCnet Using Forehead iPPG for Continuous and Noncontact Blood Pressure Measurement

Abstract

Camera-based image sequences can be used excitingly to provide noncontact and continuous blood pressure (BP) measurement. This study proposes an innovative deep-learning approach that achieves the performance of current state of the art for BP detection using forehead imaging photoplethysmography (iPPG). We studied 814 subjects with iPPG waveforms and utilize a hybrid deep 1-D convolutional neural network (Hybrid D1DCnet) to predict their systolic BP (SBP) and diastolic BP (DBP). Our model uses deep 1-D convolutions to extract high-level morphological information of waveforms, uses the attention-based long short-term memory (LSTM) unit to learn temporal-dependent features, and combines personal information through the multilayer perceptron (MLP) to perform regression estimation. This model is capable of continuously monitoring BP by detecting noncontact facial iPPG signal over a short-time period of 8 s. The mean absolute error (MAE) and the standard deviation (STD) of estimated SBP and DBP are 8.36 ± 6.22 and 5.69 ± 3.97 mmHg, respectively. Interpretability studies show that the middle part of the rising edge of the iPPG pulse waveform is of most interest. It is also verified that this model has interdevice generalizability. This scheme achieves the C/B grade (systolic/diastolic) according to the British Society of Hypertension grading criteria. In the current work, we demonstrate the feasibility and advantages of the Hybrid D1DCnet-based noncontact and continuous BP measurement using forehead iPPG waveforms. Such applications of monitoring circulatory health are expected to promote the prevention and intervention of cardiovascular diseases for the public.