Exploiting Multiwavelength Morphological Features of Camera-PPG for Blood Pressure Estimation

The feasibility of noncontact blood pressure (BP) estimation using remote photoplethysmography (rPPG) has been demonstrated, and its application can reduce the workload of caregivers and enhance patient comfort in long-term monitoring. However, the existing studies in rPPG waveform morphology-based BP regression ignore the wavelength dependence of rPPG signals, where signals from different wavelengths vary in morphology. It may lead to performance variations in BP estimation, which should be explored to increase our understanding of camera-based BP estimation. In this study, we investigate multiwavelength morphological features of rPPG for BP estimation, where the rPPG signals are extracted from both the visible and near-infrared (NIR) wavelength, and a deep neural network is customized to learn its morphological features for BP estimation. We created a dataset including two groups of subjects for benchmarking: healthy participants from the laboratory and critically-ill patients from the hospital intensive care unit (ICU). The subject-independent benchmark results show that although the rPPG signals from green (G) wavelength have stronger pulsatile strength, the ones from red (R) and NIR wavelengths with deeper penetration depth exhibit superior performance in BP estimation. It also shows that, compared with BP estimation at a single wavelength and other pairwise, the combination of R and NIR wavelengths can further reduce the standard deviation of both the mean absolute error (MAE) and the root-mean-square error (RMSE). Specifically, the MAE is reduced by 0.591–3.997mmHg for systolic BP (SBP), 0.318–1.133mmHg for diastolic BP (DBP), and 0.724–2.007mmHg for mean BP (MBP). The RMSE is reduced by 0.629–3.983mmHg for SBP, 0.268–1.111mmHg for DBP, and 0.643–1.893mmHg for MBP. It suggests that the fusion of rPPG morphological features at different wavelengths enables more stable camera-based BP estimation.