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

Abstract: Continuous oxygen saturation (SpO$_2$) estimation from wearable photoplethysmography (PPG) is important for long-term health monitoring, but low-quality red and infrared PPG segments can distort waveform morphology and degrade SpO$_2$ prediction accuracy. Existing PPG denoising and reconstruction methods usually optimize waveform fidelity or heart rate characteristics, while time-domain waveform loss on PPG signals alone insufficiently preserves frequency structure and SpO$_2$-relevant information. This paper proposes a SpO$_2$ predictor-guided stage-wise time-frequency reconstruction framework for low-quality dual-wavelength PPG signals. The proposed method first selects high-quality PPG segments to pretrain a SpO$_2$ predictor. A masked reconstruction model is then trained to recover randomly masked PPG regions using a joint reconstruction objective that combines time-domain waveform loss with frequency-domain loss computed from the short-time Fourier transform (STFT). To make the reconstruction task physiologically relevant, the pretrained SpO$_2$ predictor is incorporated as an additional constraint, encouraging the reconstructed PPG to preserve SpO$_2$ information rather than only minimizing waveform reconstruction error. The SpO$_2$ predictor and PPG reconstructor model are optimized through four training stages. Experiments on the public OpenOximetry Repository and a private wearable PPG dataset show that the proposed approach achieves the lowest subject-level MAE, with 2.882\% on the public dataset and 2.359\% on the private dataset.

• The paper introduces a four-stage framework that combines a frozen SpO₂ predictor with Transformer-based masked reconstruction and joint time- and STFT-frequency losses for degraded dual-wavelength PPG.
• The method achieves 2.882% subject-level MAE on OpenOximetry and 2.359% on a private wristband dataset, outperforming direct prediction and conventional ratio-based methods.
• Ablations show predictor guidance is the most important component, while unguided reconstruction can reduce SpO₂ accuracy below the level achieved by skipping reconstruction entirely.

Overview

This paper addresses a persistent problem in wearable pulse oximetry: low-quality red and infrared PPG segments distort waveform morphology and degrade continuous SpO$_2$ estimation. The authors propose a framework in which PPG reconstruction is explicitly aligned with the downstream physiological task rather than treated as generic denoising. The central idea is to use a pretrained SpO$_2$ predictor as a frozen physiological constraint during masked reconstruction of dual-wavelength PPG, combined with a joint time-domain and STFT-based frequency-domain reconstruction objective, optimized through four alternating training stages. On the public OpenOximetry Repository the framework achieves a subject-level MAE of 2.882% (RMSE 4.297), improving over a direct-prediction Bi-LSTM attention baseline at 3.063% MAE; on a private We-Be wristband dataset it reaches a subject-level MAE of 2.359%.

The paper's motivating claim is that time-domain waveform losses alone insufficiently preserve SpO$_2$-relevant information — a claim the ablation results support: removing the predictor-guided loss degrades performance more than removing either signal-fidelity loss, and under longer mask durations the variant without predictor guidance performs worse than simply skipping reconstruction altogether.

Method

Preprocessing and quality-based segment selection

Signals are band-pass split into AC (0.5–5 Hz) and DC (<0.5 Hz) components; the normalized AC/DC ratio is segmented into 10-s windows with 1-s sliding steps. High-quality training segments are retained via template-matching SQA (NeuroKit2 implementation of the Orphanidou method), requiring all 1-s red-channel quality scores ≥ 0.6. All splits are subject-level.

Four-stage optimization

• Stage 1: A Bi-LSTM with attention SpO$_2$ predictor is pretrained with MSE loss on high-quality segments.
• Stage 2: A four-layer Transformer encoder reconstructor learns to fill randomly zero-masked contiguous regions of 1–5 s. Its objective combines three terms:

$$\mathcal{L}_{\mathrm{recon}} = \mathcal{L}_{\mathrm{time}} + \lambda_{\mathrm{freq}}\mathcal{L}_{\mathrm{freq}} + \lambda_{\mathrm{SpO}_2}\mathcal{L}_{\mathrm{SpO}_2}^{\mathrm{guide}}$$

where $\mathcal{L}_{\mathrm{time}}$ is MSE over masked samples, $\mathcal{L}_{\mathrm{freq}}$ is MSE over STFT frames overlapping the mask (FFT size 200, hop 20, complex components of both channels), and $\mathcal{L}_{\mathrm{SpO}_2}^{\mathrm{guide}}$ passes the merged reconstructed signal through the frozen pretrained predictor against ground-truth SpO$_2$. Loss weights are set empirically so weighted terms remain of the same order of magnitude.

• Stage 3: With the reconstructor frozen, the SpO$_2$ predictor is refined on reconstructed representations from all segments (high- and low-quality), masking each segment's lowest-quality $_2$0 s region. This adapts the predictor to the reconstructor's output distribution.
• Stage 4: With the refined predictor frozen, the reconstructor is further fine-tuned on high-quality segments using the same Stage-2 objective.

At test time, every segment — regardless of quality — has its lowest-quality 3-s region masked, reconstructed, merged, and passed through the final predictor.

Results

Ablations

On OpenOximetry, the full model attains subject-level MAE/RMSE of 2.882/4.297 versus 3.071/4.483 without the guided loss, 2.935/4.320 without the frequency loss, and 2.948/4.438 without the time-domain loss. The private dataset shows the same ordering (full model 2.359 vs. 2.522 without guidance). Two observations stand out. First, the predictor-guided term is the most critical component by a consistent margin across both datasets and both granularities. Second, the full model achieves the best frequency-domain reconstruction MAE for both channels (0.00446 red, 0.00531 IR) while accepting a marginal increase in time-domain MAE relative to the variant trained only on fidelity losses (0.00223 vs. 0.00218 red) — direct evidence that the guided loss trades a small amount of waveform fidelity for downstream accuracy.

One inconsistency deserves note: the full model does not achieve the best segment-level RMSE on OpenOximetry (4.913 vs. 4.911 without the time-domain loss), a minor concession within an otherwise consistent ablation picture.

Comparisons and stage-wise analysis

Against related methods at matched preprocessing, the proposed approach outperforms conventional ratio-of-ratios calibration (MAE 3.460), NormWear foundation-model embeddings (3.390), and the direct Bi-LSTM baseline (3.063). Stage-wise results show monotonic improvement: Stage 2 alone yields 3.370 MAE, Stage 3 refines this to 2.968, and Stage 4 reaches 2.882, indicating that alternating optimization between the two modules contributes materially beyond either component alone.

Mask duration and signal quality

Mask duration analysis reveals an empirical trade-off: errors for the full model first decrease then slightly increase with longer masks, optimal at $_2$1 s, whereas ablated variants degrade monotonically. Critically, the variant lacking predictor-guided loss performs worse than no reconstruction at all at several mask durations — a strong indication that unconstrained reconstruction can destroy SpO₂ information. Signal-quality binning shows the benefit concentrates where intended: below quality score 0.6, guided variants beat the baseline predictor, although most test segments lie in high-quality ranges, which limits how much of the aggregate improvement is attributable to low-quality handling. Case visualizations suggest a denoising effect on distorted masked regions, though this remains qualitative.

Limitations and open questions

Several limitations are acknowledged or evident. The private wearable study covers only 9 subjects evaluated leave-one-subject-out, so its 2.359% MAE carries substantial uncertainty. The loss weights ($_2$2, $_2$3) are tuned empirically by monitoring loss magnitudes rather than via principled selection or sensitivity analysis. The fixed $_2$4 s masking strategy applied uniformly at test time — even to already-high-quality segments — is suboptimal, and the authors themselves identify adaptive mask placement and duration as open work. Additionally, because reference SpO$_2$5 labels are used to supervise reconstruction indirectly through the guided loss, the framework presupposes reliable labels aligned to degraded segments; performance under label noise is not examined.

Conclusion

This work reframes PPG reconstruction as downstream-task-aligned restoration: by coupling a frozen SpO$_2$6 predictor into a masked Transformer reconstructor's objective alongside complementary time- and frequency-domain losses, and by alternately refining both modules over four stages, the authors obtain consistent gains over direct prediction and prior baselines on public and private data. The evidence that unguided reconstruction can be worse than no reconstruction is arguably the most instructive finding. Remaining questions include whether adaptive masking generalizes across devices and whether the guided-reconstruction paradigm transfers to other PPG tasks such as heart-rate and respiratory estimation.