Andrii Bidochko

• Updated: July 14, 2026
• 2 min read

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

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

Authors: Zequan Liang, Elahe Hosseini, Ning Miao, Mahdi Pirayesh Shirazi Nejad, Wei Shao, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun

Published: July 11, 2026

Abstract

Continuous oxygen saturation (SpO₂) estimation from wearable photoplethysmography (PPG) is essential for long‑term health monitoring. Low‑quality red and infrared PPG segments distort waveform morphology and degrade SpO₂ prediction accuracy. Existing denoising methods focus on waveform fidelity or heart‑rate characteristics, neglecting the frequency structure that carries crucial SpO₂ information. This paper introduces a SpO₂ predictor‑guided stage‑wise time‑frequency reconstruction framework that jointly optimises time‑domain waveform loss and frequency‑domain loss (via short‑time Fourier transform) while enforcing a pretrained SpO₂ predictor as a physiological constraint.

Key Contributions

• Pre‑training of a high‑quality PPG‑based SpO₂ predictor to serve as a reconstruction constraint.
• Masked reconstruction model that recovers randomly masked PPG regions using a combined time‑ and frequency‑domain loss.
• Four‑stage training pipeline that aligns waveform fidelity with SpO₂‑relevant information.
• State‑of‑the‑art performance on the OpenOximetry Repository (MAE = 2.882 %) and a private wearable dataset (MAE = 2.359 %).

Methodology

The framework operates in three phases:

1. High‑quality segment selection: Identify clean dual‑wavelength PPG segments to train the SpO₂ predictor.
2. Masked reconstruction training: Randomly mask portions of low‑quality PPG and train a reconstruction network using a joint loss:

   Loss = λ₁·‖x̂‑x‖₂ + λ₂·‖STFT(x̂)‑STFT(x)‖₂ + λ₃·‖Predictor(x̂)‑SpO₂‖₂

3. Physiological constraint integration: The pretrained predictor guides the reconstruction to preserve SpO₂‑relevant features.

Results

Extensive experiments demonstrate that the proposed method consistently outperforms baseline denoising and reconstruction approaches. The MAE improvements translate to more reliable SpO₂ monitoring in real‑world wearable scenarios.

Conclusion

By coupling a SpO₂ predictor with a stage‑wise time‑frequency reconstruction strategy, we achieve superior preservation of both waveform morphology and the underlying physiological signal. This work paves the way for robust, low‑cost wearable SpO₂ monitoring solutions.

For more details on our research and related projects, visit the UBOS Research Hub or explore our technology blog.

References

Liang, Z., Hosseini, E., Miao, N., et al. (2026). SpO₂ Predictor‑Guided Stage‑Wise Time‑Frequency Reconstruction of Low‑Quality Dual‑Wavelength PPG for Oxygen Saturation Estimation. arXiv preprint arXiv:2607.07996. https://arxiv.org/abs/2607.07996

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.