OxyMamba: Multi-Temporal Parallel Mamba for Camera-Based Blood Oxygen Estimation.

Abstract

Blood oxygen saturation (SpO$_{2}$) is a critical physiological parameter for evaluating respiratory and circulatory health. Current contactless video-based SpO$_{2}$ estimation methods still face three key challenges: limited modeling of long temporal dependencies, insufficient utilization of complementary RGB spectral information, and limited incorporation of physiological priors. To address these issues, we propose OxyMamba, a state-space-model-based framework that combines (1) a Multi-Temporal Parallel Mamba module for long temporal context modeling, (2) a cross-channel complex-domain mixing module for RGB amplitude-phase interaction modeling in the frequency domain, and (3) an anchor-based heart-rate sideband gating mechanism for emphasizing physiologically plausible spectral components. Under the evaluated dataset-specific protocols, OxyMamba achieves MAE/RMSE values of 0.49%/0.62% on PURE, 1.14%/1.45% on VIPL-HR, and 0.89%/1.14% on ViInHealth. Additional low-SpO$_{2}$ subgroup analysis shows that the gating mechanism improves performance over the no-gating variant within the evaluated setting. These results indicate favorable performance in the tested protocols, while broader validation across populations, acquisition conditions, skin-tone groups, and lower SpO$_{2}$ ranges is still needed before clinical-use claims can be made.