Towards generalizable contactless oximetry from multispectral video via deep domain-adaptive learning strategy

Camera-based contactless oxygen saturation estimation enables comfort and hygienic long-term monitoring. Deep learning methods have shown promise in extracting rich spatiotemporal features from facial videos, yet they often struggle to generalize across domain shifts, such as inter-subject variability, skin tone differences, spectral discrepancies, sensor system changes, and lab-to-clinic transitions. To address this, we propose a multi-level target unsupervised domain adaptation framework for contactless oximetry from multispectral facial video. Building upon a previously established spatiotemporal 3D CNN baseline, the proposed method aligns source and target feature distributions across multiple network hierarchies together with a dynamically weighted training scheduler and adaptive batch normalization. The framework is systematically evaluated on three custom datasets acquired with two different sensor systems, covering controlled laboratory experiments and real-world clinical recordings. In leave-one-participant-out validation on a breath-holding dataset of 23 healthy subjects, the proposed method reduces the mean absolute error (MAE) from 2.31 % to 1.98 %, desaturation-specific MAE from 3.26 % to 2.60 %, and improves Pearson's correlation coefficient from 0.64 to 0.71. Consistent performance gains are observed under cross-skin-type, cross-spectral, and cross-sensor-system domain shifts. In a clinical validation involving real sleep apnea patients with 796 oxygen desaturation events, the error between our estimations and polysomnography ground truth stays within 2 % for 90 % of the recorded time. These results demonstrate that the proposed deep domain adaptation framework substantially enhances the robustness and generalization of camera-based contactless oximetry, demonstrating its potential for practical deployment in clinical and real-world monitoring.