⚡ electrical engineering

SpO $_2$ Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG for Oxygen Saturation Estimation

This paper proposes a SpO $_2$ predictor-guided stage-wise time-frequency reconstruction framework that combines time-domain waveform loss with frequency-domain constraints to effectively recover low-quality dual-wavelength PPG signals, thereby achieving superior oxygen saturation estimation accuracy compared to existing methods.

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

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your smartwatch is trying to tell you how much oxygen is in your blood, but your hand is shaking, or the strap is loose, or you're just moving around too much. The signal it gets—the "PPG" wave that looks like a heartbeat on a screen—gets all squished, noisy, and distorted. It's like trying to listen to a favorite song while someone is banging on a drum next to you; you can hear the beat, but the melody is a mess.

For a long time, scientists tried to fix this by just cleaning up the noise, like using a filter to make the song sound clearer. But the authors of this paper argue that simply making the waveform look "clean" isn't enough. It's like fixing a broken radio to make the static go away, but if the station you're tuned to is still playing the wrong song, you haven't really solved the problem. The paper explicitly argues against methods that only focus on making the signal look perfect in the time domain (the shape of the wave) or just preserving the heart rate, because those methods often lose the specific "flavor" needed to calculate oxygen levels accurately.

So, what did they do? They built a clever, four-step training system that acts like a master chef teaching an apprentice how to cook a specific dish, rather than just teaching them how to chop vegetables.

The Four-Step Cooking Class

The Taste Test (Pretraining): First, they take a bunch of perfect, high-quality signals (the "fresh ingredients") and train a smart computer brain (an SpO2 predictor) to recognize exactly what a healthy oxygen level looks like. This brain learns the "taste" of the data.
The Blindfolded Chef (Masked Reconstruction): Next, they take a perfect signal and randomly hide a chunk of it—like covering a section of a song with a piece of tape. They then train a second model (the reconstructor) to guess what's under the tape. But here's the twist: they don't just want the chef to guess the shape of the wave. They use the "Taste Test" brain from step one as a strict supervisor. If the chef guesses a shape that looks okay but sounds like the wrong song (wrong oxygen level), the supervisor yells, "No! Try again!" This forces the model to learn not just to fill in the gaps, but to fill them with the right kind of information.
The Master's Refinement: Once the chef gets good at guessing, they let the "Taste Test" brain learn from the chef's new, reconstructed signals. This helps the brain get better at reading even the messy, noisy signals.
The Final Polish: Finally, they go back and refine the chef one last time, using the now-smarter brain to make sure every reconstructed signal is perfect for the final goal.

The Secret Sauce: Time and Frequency

The paper suggests that to fix these signals, you have to look at them in two ways at once. Imagine a song: you need to hear the notes (the time domain) and also see the sheet music (the frequency domain). The authors combined a loss function that checks the shape of the wave with another that checks the "spectrum" or frequency structure. They found that if you only check the shape, you miss the frequency details; if you only check the frequency, you miss the shape. You need both.

Did it Work?

The authors tested this on two different sets of data: a public collection of signals called the OpenOximetry Repository and a private dataset from a wearable band they made themselves.

On the public dataset, their method achieved a subject-level Mean Absolute Error (MAE) of 2.882%. On the private dataset, it hit 2.359%. To put that in perspective, they compared their method to a standard "Calibration" method (which got 3.460% error) and a "Baseline" AI model (which got 3.063% error). Their approach was the most accurate of all the methods they tested.

The paper suggests that the most critical part of their success was that "Taste Test" supervisor. When they removed the part of the system that used the SpO2 predictor to guide the reconstruction, the performance got worse than even the basic AI model. This suggests that forcing the reconstruction to care about the final oxygen number is the key to making it work.

They also looked at what happens when they hide different amounts of the signal (from 1 second to 5 seconds). They found that hiding 3 seconds of the signal (k = 3 s) seemed to be the sweet spot for their setup. Interestingly, even when the original signal was very noisy (quality scores below 0.6), their method kept the error lower than the baseline, suggesting it really does help preserve the oxygen information even when the data is bad.

In short, the paper doesn't claim to have solved the problem of noisy sensors forever, but it suggests that by training a signal-repair model to care about the meaning of the signal (the oxygen level) rather than just the look of the signal, we can get much more accurate health readings from our wobbly, moving wrists.

Technical Summary: SpO2 Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG

Problem Statement
Continuous oxygen saturation ( $SpO_2$ ) estimation from wearable photoplethysmography (PPG) is critical for long-term health monitoring. However, wearable acquisition is frequently compromised by motion artifacts, unstable skin contact, and optical coupling variations. These factors distort the morphology and frequency characteristics of red and infrared PPG signals, degrading downstream $SpO_2$ prediction accuracy. Existing denoising and reconstruction methods typically optimize for time-domain waveform fidelity or heart rate characteristics. The authors argue that minimizing time-domain loss alone is insufficient because it fails to preserve the frequency structure and $SpO_2$ -relevant information necessary for accurate physiological estimation. Furthermore, supervised reconstruction often requires clean reference signals, which are difficult to obtain when source signals are already degraded.

Methodology
The paper proposes a SpO2 predictor-guided stage-wise time-frequency reconstruction framework designed to recover low-quality dual-wavelength PPG signals while preserving information relevant to oxygen saturation. The framework operates through four distinct training stages:

PPG Preprocessing: Raw red and infrared PPG signals are decomposed into AC (0.5–5 Hz) and DC (<0.5 Hz) components. The normalized signal (AC/DC) is segmented into 10-second windows.
Stage 1: SpO2 Predictor Pretraining: A Bi-LSTM with attention is pretrained on high-quality PPG segments (selected via signal quality assessment) to serve as a reliable $SpO_2$ estimator.
Stage 2: Predictor-Guided PPG Reconstructor Training: A Transformer-based reconstructor is trained to recover randomly masked temporal regions (1–5 seconds) from high-quality segments. The model is optimized using a joint reconstruction objective comprising:
- Time-Domain Loss ( $L_{time}$ ): Minimizes the difference between reconstructed and original waveforms in the masked region.
- Frequency-Domain Loss ( $L_{freq}$ ): Minimizes the difference between the Short-Time Fourier Transform (STFT) of the reconstructed and original signals, preserving spectral structure.
- Predictor-Guided $SpO_2$ Loss ( $L_{guide}^{SpO_2}$ ): The frozen pretrained $SpO_2$ predictor acts as a physiological constraint. The loss is calculated based on the $SpO_2$ estimation error of the reconstructed signal (merged with unmasked regions), encouraging the reconstructor to preserve $SpO_2$ -relevant morphology rather than just minimizing waveform error.
Stage 3: SpO2 Predictor Refinement: The reconstructor is frozen, and the $SpO_2$ predictor is further trained on the entire dataset (including low-quality segments processed by the reconstructor) to adapt to reconstructed representations.
Stage 4: PPG Reconstructor Refinement: The refined predictor is frozen, and the reconstructor is further optimized on high-quality segments using the joint objective from Stage 2, leveraging the improved physiological guidance.

During inference, the lowest quality 3-second segment of a test signal is masked, reconstructed, merged, and fed into the final predictor.

Key Contributions

Physiological Constraint in Reconstruction: The authors introduce a strategy where a pretrained $SpO_2$ predictor guides the PPG reconstruction process. This ensures the reconstructed signals retain physiological information critical for the downstream task, rather than merely optimizing for generic signal fidelity.
Joint Time-Frequency Objective: The framework combines time-domain waveform loss with frequency-domain loss (via STFT) to ensure the reconstruction preserves both morphological trends and spectral characteristics.
Stage-Wise Optimization: A four-stage training strategy is designed to alternately refine the $SpO_2$ predictor and the PPG reconstructor, enabling joint improvement of signal restoration and estimation accuracy.

Experimental Results
The framework was evaluated on the public OpenOximetry Repository (191 subjects) and a private wearable PPG dataset (9 subjects).

Performance: The proposed method achieved the lowest subject-level Mean Absolute Error (MAE) among all compared methods.
- OpenOximetry Dataset: 2.882% MAE (compared to 3.063% for a baseline Bi-LSTM and 3.460% for conventional calibration).
- Private Wearable Dataset: 2.359% MAE.
Ablation Studies: Removing the predictor-guided $SpO_2$ loss resulted in the most significant performance degradation, confirming its critical role. Removing either time-domain or frequency-domain losses also reduced performance, indicating that both waveform fidelity and spectral consistency are necessary.
Robustness: The method demonstrated improved performance on low-quality PPG segments (quality scores < 0.6) compared to baseline predictors, suggesting the reconstruction effectively denoises and recovers $SpO_2$ information.

Significance
The paper claims that treating PPG reconstruction as a generic denoising problem is suboptimal for $SpO_2$ estimation. By aligning the reconstruction objective with the downstream physiological task through a predictor-guided constraint, the proposed framework achieves superior estimation accuracy. The results suggest that preserving $SpO_2$ -relevant information during signal recovery is more effective than simply minimizing reconstruction error, offering a robust solution for wearable health monitoring in the presence of motion artifacts and signal degradation.

Technical Summary: SpO2 Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG

More like this