Real-Time Subject-Specific Predictive Modeling of PPG Signals for Artifact-Resilient SpO2 Estimation Under Hypoxia

Photoplethysmography (PPG) is widely used in health monitoring, but its reliability is often compromised by artifacts, limiting accurate peripheral arterial oxygen saturation (SpO2) estimation. Moreover, physiological and demographic factors can substantially alter PPG waveform morphology. We propose a lightweight, real-time predictive modeling approach that adapts to subject-specific PPG signal dynamics to improve monitoring robustness under conditions prone to artifacts. A total of 459 min of dual-wavelength PPG signals, together with reference SpO2 values, were collected from 17 healthy volunteers (2 female, 15 male, mean age 27±3 years old) undergoing controlled desaturation in the 85-100% range after being instructed to remain still. Cardiac pulses were segmented and decomposed into AC and DC components, and the adequacy of several signal models, ranging from sums of Gaussians to Fourier series, and polynomial expansions of different orders, was evaluated. A space of representative signal features was built from the best-performing model, and used to generate machine learning-based predictions for each pulse using the preceding four clean pulses. Predicted pulses could be directly compared with their originals, enabling accurate error estimation without simulated data. The predicted signals closely matched the originals, achieving mean R2 scores above 0.9, and an SpO2 estimation RMSE of 1.28%. In practical use, the same approach could be applied to overcome artifact-corrupted segments if combined with a signal quality assessment module. Therefore, this algorithm provides a promising pathway toward more reliable SpO2 monitoring in wearable systems, particularly under hypoxic conditions.