Contactless Human Oxygen Saturation Detection Using RGB Camera in an Induced Hypoxemia Study with Varied Skin Types

Abstract Contactless monitoring of blood oxygen saturation (SpO2) using consumer-grade cameras could enable continuous monitoring in clinical settings where contact sensors are impractical, yet existing methods require optical filters, skin contact, or are restricted to narrow saturation ranges. We developed a physics-based framework that extracts ratio-of-ratios (RoR) features from facial video and predicts SpO2 via a RF regression model, evaluated through leave-one-subject-out cross-validation (LOSOCV) on 19 participants (Fitzpatrick types II–VI) during controlled induced hypoxemia (SpO2: 70–100%) with an unmodified commercial RGB camera at approximately 0.6-meter (2-foot) distance under ambient lighting. LOSOCV yielded ARMS of 6.62% (95% CI: 5.60–7.93%; PCC = 0.595), reducible to 3.87% through per-subject calibration. Slope-constrained one-point calibration using a single concurrent pulse oximeter reading achieved ARMS of 4.20%. Within clinically defined SpO2 bins, ARMS ranged from 2.58% to 2.73%, with error concentrating at range boundaries rather than within clinically actionable bands. For hypoxemia screening at the 90% threshold, the area under the receiver operating characteristic curve (AUC-ROC) was 0.797 (95% CI: 0.669–0.870) with sensitivity 0.636 and specificity 0.846. Performance for Fitzpatrick type VI was substantially degraded (ARMS = 9.71%), consistent with melanin-induced signal attenuation in the visible spectrum. These results demonstrate the feasibility of contactless SpO2 estimation from unmodified RGB cameras and identify skin pigmentation as the primary barrier to equitable performance.