rPPG-Guided Multi-Task Cascaded Learning for Joint Heart Rate and SpO₂ Estimation

This is the authors' abstract. We don't add key points for this paper.

Remote physiological sensing from facial videos enables contact-free health monitoring. Despite remarkable advances in deep learning-based rPPG and heart rate estimation, video-based SpO2 measurement remains challenging due to weak RGB spectral cues and the absence of physiologically grounded modeling. Existing approaches usually predict SpO2 separately and overlook its inherent correlation with rPPG-related pulsatile blood volume dynamics, which degrades SpO2 accuracy as well as rPPG robustness under real-world conditions. In this paper, we propose a novel rPPG-Guided Multi-Task Cascaded Learning framework to jointly recover rPPG waveforms and estimate SpO2 from facial videos. The model explicitly captures the physiological coupling between pulse dynamics and blood oxygen saturation within a unified end-to-end architecture. Extensive experiments on public benchmarks demonstrate that the proposed framework achieves competitive overall performance and yields improved SpO2 prediction accuracy and stable rPPG reconstruction on most test scenarios. These results verify the efficacy of physiology-aware cascaded modeling in reducing estimation errors for video-based vital sign measurement.