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

Area of Science:

• Biomedical Engineering
• Signal Processing
• Wearable Health Technology

Background:

• Photoplethysmography (PPG) is crucial for health monitoring, but artifacts and physiological variations reduce peripheral arterial oxygen saturation (SpO2) accuracy.
• Existing methods struggle with signal degradation, limiting reliable SpO2 estimation in real-world applications.

Purpose of the Study:

• To develop a lightweight, real-time predictive modeling approach for robust PPG-based SpO2 monitoring.
• To adapt signal processing to subject-specific PPG dynamics and mitigate artifact impact.

Main Methods:

• Collected 459 minutes of dual-wavelength PPG signals and reference SpO2 from 17 volunteers during controlled desaturation.
• Segmented and decomposed cardiac pulses, evaluating various signal models (Gaussians, Fourier series, polynomials).
• Built a feature space from the best model to generate machine learning predictions using preceding clean pulses.

Main Results:

• Predicted PPG signals closely matched original signals, achieving mean R2 scores above 0.9.
• Achieved a low SpO2 estimation Root Mean Square Error (RMSE) of 1.28%.
• Demonstrated accurate error estimation by comparing predicted vs. original pulses without simulated data.

Conclusions:

• The proposed algorithm offers a promising solution for reliable SpO2 monitoring in wearable systems, especially in challenging conditions.
• Integration with signal quality assessment can enable overcoming artifact-corrupted segments for practical use.
• This adaptive, predictive approach enhances PPG signal robustness against artifacts and individual variability.