SpO2 Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG for Oxygen Saturation Estimation

Continuous monitoring of oxygen saturation (SpO2) using wearable photoplethysmography (PPG) is frequently hindered by motion artifacts and poor sensor contact, which distort signal morphology and degrade estimation accuracy. The authors address the challenge of restoring these degraded dual-wavelength (red and infrared) PPG signals to ensure reliable SpO2 estimation, moving beyond simple denoising to preserve physiologically relevant information.

The researchers introduce a four-stage, predictor-guided reconstruction framework. First, a Bi-LSTM model is pretrained on high-quality PPG segments to serve as an SpO2 estimator. Second, a transformer-based masked autoencoder is trained to recover missing or corrupted temporal regions of PPG signals. This reconstructor is optimized using a joint objective function that minimizes time-domain waveform errors, frequency-domain spectral discrepancies (via STFT), and a physiological loss term derived from the frozen SpO2 predictor. The final two stages involve iterative refinement of the predictor and the reconstructor to ensure the recovered signals are optimized for the downstream task of SpO2 estimation.

The proposed method significantly improves SpO2 estimation accuracy compared to baseline approaches. Experiments on the public OpenOximetry Repository and a private wearable dataset demonstrate that the joint time-frequency reconstruction, guided by the SpO2 predictor, achieves the lowest subject-level Mean Absolute Error (MAE) of 2.882% and 2.359%, respectively. Ablation studies confirm that the SpO2 predictor-guided loss is the most critical component for maintaining physiological accuracy, while the combination of time and frequency domain losses ensures robust signal restoration.