Oxygen Saturation Estimation Based on Optimal Band Selection From Multi-Band Video

Ryo Takahashi, Koichi Ashida, Yasuo Kobayashi, Rumi Tokunaga, Shuhei Kodama, Norimichi TsumuraWorkshops @ CVPR 2021 · DOI

Abstract

Researchers have developed a method to improve non-contact oxygen saturation ($SpO_2$) monitoring by dynamically selecting the best two spectral bands from multi-band video. By using Monte Carlo simulations to model light-skin interactions, they trained a regression model that identifies the most reliable band combination based on real-time signal features, potentially making remote health monitoring much more robust across different environments and skin types.

Introduction

Measuring oxygen saturation ($SpO_2$) is critical for managing respiratory conditions, a need highlighted globally by the COVID-19 pandemic. While traditional pulse oximeters are accurate, they are uncomfortable for long-term use. Non-contact, camera-based methods typically use the Red and Blue channels of a standard RGB camera to calculate the "Ratio of Ratios" (RoR)—a value that correlates with blood oxygen levels.

Absorption coefficients of oxidized and deoxidized hemoglobin across the visible and near-infrared spectrum.

As smartphones and surveillance systems adopt multi-band cameras (e.g., RGB plus multiple Infrared bands), a new challenge arises: identifying which two bands to use. The "best" pair changes depending on the lighting conditions and the subject's skin characteristics. Using the wrong pair leads to noisy data and inaccurate health readings.

Methodology

The proposed framework selects the optimal band pair automatically through two main phases:

  1. Simulation via Monte Carlo Modeling: Due to the difficulty of collecting vast amounts of real-world multi-band video with varying $SpO_2$ levels, the team used MCML (Monte Carlo model of light transport in multi-layered tissues). They simulated light interactions with skin by varying melanin levels, hemoglobin concentration, and oxygen saturation (80%–100%), generating 51 spectral patterns and over 62,000 pulse wave variations.
  2. Feature Extraction and Selection: Three key features determine if a band pair is "good":
  3. SNR (Signal-to-Noise Ratio): Calculated for both candidate bands.
  4. Normalized RoR Difference ($d\hat{RoR}$): A novel feature representing how much the ratio changes as $SpO_2$ drops. If this difference is too small, the signal is easily lost to noise.
Relationship between SpO2 levels, pulse wave amplitudes in Red and IR bands, and the resulting Ratio of Ratios (RoR).
  1. Regression Model: A Support Vector Regression (SVR) model with a Radial Basis Function (RBF) kernel was trained to predict the correlation coefficient between the calculated RoR and the actual $SpO_2$. The SVR acts as a "judge" to predict which band pair will yield the most accurate results.

Results and Discussion

The researchers validated their SVR model using a test set of simulated data. The best-performing model achieved a Coefficient of Determination ($R^2$) of 0.8664 and a Mean Absolute Error (MAE) of 0.1329 for predicting the correlation coefficient. Optimization through grid search identified specific hyperparameters ($C=1, \gamma=1$) that provided the best balance for the regression model.

A key insight from this study is the "Similarity Trap": if two spectral bands are too close to each other, their absorption characteristics become nearly identical. This causes the RoR to stay flat even as oxygen levels change, making the signal useless. The $d\hat{RoR}$ feature was specifically designed to detect and avoid this issue. Furthermore, the paper demonstrates that high-fidelity physics simulations (MCML) can effectively stand in for real-world data to train machine learning models for physiological sensing.

Conclusion

This research presents a dynamic adaptation system that evaluates the environment and the subject to pick the best spectral tools for $SpO_2$ estimation in real-time. While the study is currently simulation-based, it provides a robust foundation for future work. Next steps include real-world validation with actual multi-band camera hardware, modeling more complex environmental noise, and integrating the algorithm into mobile devices for low-latency health monitoring.

Summary updated

BibTeX
@inproceedings{takahashi2021oxygen,
  author = {Ryo Takahashi and Koichi Ashida and Yasuo Kobayashi and Rumi Tokunaga and Shuhei Kodama and Norimichi Tsumura},
  title = {{Oxygen Saturation Estimation Based on Optimal Band Selection From Multi-Band Video}},
  booktitle = {Workshops @ CVPR},
  year = {2021},
  doi = {10.1109/cvprw53098.2021.00426},
}
Figure