Video-based SpO₂ estimation in multi-scenarios via wavelength-guided image feature enhancement

Video-based physiological signal monitoring using RGB cameras has attracted increasing attention due to its convenience, low cost, and non-contact nature. Owing to the limited spectral information available in the visible spectrum and the difficulty of capturing oxygen-related cues, video-based blood oxygen saturation (SpO₂) estimation remains immature. To address the aforementioned challenges, this study builds upon the optical absorption characteristics of SpO₂-related components in the near-infrared range and identifies several visible-wavelength bands that exhibit similar absorption behaviors within the visible spectrum. To further enhance the representation of these wavelength-related characteristics in RGB images, we propose a standard colorimetric distance–guided image feature enhancement method, whose effectiveness is validated through auxiliary experiments involving spectral imaging. The enhanced spatiotemporal representations are then fed into a lightweight deep neural network for accurate SpO₂ estimation. We construct a real-world ICU SpO₂ dataset, providing a valuable benchmark for algorithm development and evaluation. Experimental results demonstrate that the proposed method achieves superior performance on both public and in-house datasets.