Non-contact Blood Pressure Estimation with Multi-spectral Imaging Photoplethysmogram

Imaging photoplethysmography (iPPG) acquired via cameras (RGB, infrared, etc.) holds great promise for vital sign monitoring due to its non-contact and low-cost advantages. Although RGB cameras are ubiquitous in daily life (e.g., smartphones, laptops, surveillance cameras) and have been utilized for blood pressure (BP) estimation, most existing studies fused multi-spectral iPPG (Red, Green, Blue channels) into a single channel, neglecting the distinct cardiovascular information carried by iPPG of different spectral and from different anatomical regions. To address this gap, this study systematically explores the relationship between multi-spectral iPPGs (Red, Green, Blue) and BP, and proposes a high-accuracy non-contact BP estimation method. A dataset was constructed using 30 subjects aged 18-35 years old which have no history of cardiovascular diseases, with mean resting systolic BP (SBP) of 119 ± 12 mmHg and mean diastolic BP of 75 ± 9 mmHg. Data were collected simultaneously via an RGB camera and the CNAP monitor 500 HD under three BP states, rest, controlled deep breathing at 0.1 Hz and post-stepping exercise. Based on this dataset, we conducted two studies: the first investigated the correlation between BP and differences in iPPGs across RGB spectral within the same ROI on the forehead measuring 800×400 pixels (single-ROI). The second study examined the relationship between BP and variations in iPPGs for individual RGB spectral bands across distinct anatomical regions, between two adjacent regions on the forehead, each of size 400×200 pixels (dual-ROIs). To quantify the differences between iPPGs, following signal preprocessing, we extracted four types of differential features from the two signals being compared: the difference of the same kind of iPPG features, the difference of different kinds of iPPG features, the difference of the same iPPG fiducial points, and the difference of different kinds of iPPG fiducial points. Pearson correlation coefficient was used for difference features selection, and gradient boosting regression (GBR) is employed with LeaveOne-Out Cross-Validation (LOOCV) in this study to analyze the association between these difference features and BP. For the single-ROI model, the GBR achieved the accuracy with SBP of 0.01 ± 7.41 mmHg and DBP of 0.05 ± 4.38 mmHg. For the dualROI model, the GBR yielded SBP of -0.06 ± 7.14 mmHg and DBP of 0.02 ± 4.84 mmHg. Comparative experiments also showed that the proposed method outperformed the single-wavelength iPPG method and the fused iPPG method. This study demonstrates that multi-spectral iPPGs carry unique BP-related information, providing a novel and effective approach for non-contact BP estimation with improved performance. © 2014 IEEE.