Remote Blood Oxygen Estimation From Videos Using Neural Networks

Area of Science:

• Biomedical Engineering
• Computer Vision
• Physiological Monitoring

Background:

• Peripheral blood oxygen saturation (SpO2) is a critical health indicator, especially relevant during the COVID-19 pandemic.
• Low SpO2 can precede obvious symptoms in COVID-19 patients, necessitating accessible monitoring.
• Contact-based SpO2 measurement methods pose risks of cross-contamination and discomfort.

Purpose of the Study:

• To develop the first noncontact SpO2 estimation scheme using smartphone cameras and convolutional neural networks (CNNs).
• To provide a convenient, comfortable, and privacy-preserving method for SpO2 monitoring.
• To design explainable AI models inspired by optophysiological principles for SpO2 measurement.

Main Methods:

• Utilized smartphone videos of the hand for physiological sensing.
• Developed novel CNN architectures inspired by optophysiological models.
• Demonstrated model explainability through weight visualization for channel combination.

Main Results:

• The proposed noncontact CNN-based method outperformed state-of-the-art contact-based SpO2 estimation models.
• Analyzed the influence of skin type and hand side on SpO2 estimation accuracy.
• Validated the potential of the method for public health applications.

Conclusions:

• The developed noncontact SpO2 estimation using smartphone cameras shows significant potential for widespread public health use.
• The explainable AI approach enhances trust and understanding of the SpO2 measurement process.
• This technology offers a user-friendly and safe alternative for remote physiological monitoring.