Contactless Blood Oxygen Saturation Estimation from Facial Videos Using Deep Learning

Area of Science:

• Biomedical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Blood oxygen saturation (SpO2) is a critical health indicator.
• Traditional SpO2 measurement requires skin contact via pulse oximeters.
• Contactless SpO2 monitoring offers a significant advancement in healthcare technology.

Purpose of the Study:

• To develop and evaluate novel deep learning models for remote SpO2 measurement from facial videos.
• To assess the performance of these models against established benchmarks and conventional methods.
• To investigate factors influencing contactless SpO2 measurement accuracy.

Main Methods:

• Utilized deep learning models, specifically convolutional neural networks, for SpO2 prediction.
• Employed a spatial-temporal representation to encode SpO2 information from RGB camera videos.
• Evaluated models using the VIPL-HR public benchmark database.

Main Results:

• The best deep learning model achieved a mean absolute error of 1.274% and root mean squared error of 1.71%.
• Performance exceeded the international standard of 4% for approved pulse oximeters.
• Outperformed the conventional analytical Ratio of Ratios model for contactless SpO2 measurement.

Conclusions:

• Deep learning models can accurately measure SpO2 remotely from facial videos using standard cameras.
• Contactless SpO2 monitoring via computer vision is a viable and superior alternative to traditional methods.
• Further research insights provided on influencing factors for this emerging field.