Noncontact Multispectral SpO2 Prediction Based on Deep Ratio-of-Ratio Refinement With Optimal Band Selection and Shading Bias Removal

Noncontact prediction of peripheral oxygen saturation (SpO2) is necessary for monitoring vital signs of patients afflicted by infectious disease or sensitive to skin irritation. Recently, many approaches to imaging photoplethysmography (iPPG)-based noncontact SpO2 prediction have been proposed based on ratio-of-ratio (RoR) representing spectral ratio of diffused absorbances and deep learning (DL). Despite notable progress, issues concerning the accuracy and robustness under real-world variations and artifacts remain to be solved toward clinical viability. In this study, we proposed an approach to highly accurate and robust SpO2 prediction under significant disturbances based on multispectral iPPG signals captured from human facial regions of interest (ROIs). First, a method for accurately estimating actual diffused spectral absorbance of ROIs is proposed. To this end, the shading bias representing the discrepancy between the measured and actual diffused absorbances is calculated by projecting the measured diffused absorbance onto the actual diffused absorbance manifold preformulated based on Monte-Carlo modeling of light transport in multilayered tissues (MCML). Furthermore, the spectral band pairs optimal for estimating shading bias and actual diffused absorbances are selected, resulting in highly accurate RoRs. Second, the SpO2s derived from RoRs are spatiotemporally refined and regressed based on an LSTM AE network further to upgrade the accuracy and robustness of SpO2 prediction. Lastly, individual SpO2s predicted with multiple spectral band pairs and at multiple ROIs are fused to ensure a resilience against possible disruptions due to motion and appearance artifacts. Experiments were conducted with 12 subjects whose SpO2 levels were adjusted between 0.79-0.99 by breath-holding cycles. The cross-validated performance indicated about 2% of Mean Absolute Error (MAE) accuracy in case SpO2s were derived solely from the estimated RoRs while the accuracy of 0.46% of MAE was reached with the LSTM AE-based spatiotemporal refinement and regression.