CL-SPO2Net: Contrastive Learning Spatiotemporal Attention Network for Non-Contact Video-Based SpO2 Estimation

Area of Science:

• Biomedical Engineering
• Computer Vision
• Medical Imaging

Background:

• Non-contact peripheral oxygen saturation (SpO2) estimation using RGB cameras is a promising alternative to traditional pulse oximetry.
• Existing methods often require stable environments and large labeled datasets, limiting their practical application.
• Accuracy is frequently compromised by ambient light variations and subject motion.

Purpose of the Study:

• To develop an innovative semi-supervised network for accurate video-based SpO2 estimation.
• To overcome the limitations of small labeled datasets and environmental instability in remote photoplethysmography (rPPG) signal analysis.
• To improve the robustness and feasibility of non-contact SpO2 monitoring.

Main Methods:

• Proposed a contrastive learning spatiotemporal attention network (CL-SPO2Net) for semi-supervised video-based SpO2 estimation.
• Leveraged spatiotemporal similarities in rPPG signals from facial or hand regions.
• Integrated deep neural networks with machine learning for SpO2 calculation.

Main Results:

• Achieved a mean absolute error (MAE) of 0.85% in stable environments.
• Demonstrated robustness with MAEs of 1.13% under lighting fluctuations and 1.20% during facial rotation.
• Showcased good feasibility with small-scale labeled datasets.

Conclusions:

• CL-SPO2Net offers a viable and accurate solution for non-contact SpO2 estimation.
• The network effectively handles challenges posed by limited data and dynamic environmental conditions.
• This approach holds potential for widespread adoption in remote health monitoring.