Low-Rate Wrist SpO2 Estimation under Micro-Perturbations Using Motion-Aware Beat Selection and Perfusion-Guided Calibration

Technical Summary: Low-Rate Wrist SpO2 Estimation under Micro-Perturbations

Problem Statement
Continuous oxygen saturation (SpO2) monitoring via photoplethysmography (PPG) is critical for wearable health sensing. However, wrist-based estimation faces significant challenges compared to conventional finger-based pulse oximetry. The wrist is more susceptible to micro-perturbations, such as small movements and subtle changes in contact pressure, which destabilize the morphology of PPG beats. These perturbations alter the red and infrared pulsatile components, leading to instability in the red-to-infrared ratio-of-ratios ($R$). Since conventional methods rely on a fixed $R$–SpO2 mapping, this instability results in inconsistent estimates. Furthermore, inter-subject differences in local perfusion status introduce calibration bias, as a fixed mapping cannot adapt to varying perfusion indices (PI) across different individuals or skin-sensor contact conditions.

Methodology
The authors propose a lightweight, firmware-compatible framework for low-rate wrist SpO2 estimation that integrates motion-aware beat selection and perfusion-guided calibration. The system processes dual-wavelength (red and infrared) PPG signals alongside tri-axial accelerometer data. The methodology consists of three primary stages:

1. Beat-Level AC/DC Extraction and Ratio Computation:
   Within a 10-second sliding window, red and IR PPG signals are low-pass filtered (5 Hz cutoff) to preserve pulse morphology. A band-pass filter (1–3 Hz) facilitates robust beat detection, utilizing valleys from the more stable IR channel as common boundaries for both wavelengths. Systolic peaks are identified separately for each channel. The beat-level DC component is defined as the interpolated baseline value at the peak location, while the AC component is the amplitude difference between the peak and this baseline. The ratio-of-ratios ($R$) is computed for each beat as $R = \frac{AC_{red}/DC_{red}}{AC_{IR}/DC_{IR}}$.

2. Motion-Aware Beat Selection:
   To mitigate micro-perturbations, the framework assigns reliability weights to beats based on accelerometer-derived motion scores. For each beat interval, the standard deviation of the tri-axial accelerometer magnitude is calculated. Within a window, beats are weighted using a nonlinear function that emphasizes low-motion beats and suppresses those contaminated by motion. Specifically, a motion score threshold is established using the median and 90th percentile of the window's motion scores. A weighted median of the valid beat-level $R$ values is then computed to determine the window-level $R_{win}$, ensuring that motion-corrupted beats contribute less to the final estimate.

3. Perfusion-Guided Calibration:
   To address subject-specific variability, a reference perfusion index ($PI_{ref}$) is estimated for each session. The system identifies the most reliable 30-second segment (lowest average motion) and computes $PI_{ref}$ as the weighted median of the infrared perfusion indices from that segment. The window-level $R_{win}$ is then corrected using a linear function of $PI_{ref}$ ($R_c = R_{win} + b_0 + b_1 \cdot PI_{ref}$). Finally, the corrected ratio $R_c$ is mapped to SpO2 using a second-order polynomial calibration function learned from training data.

Key Contributions

• Low-Rate Processing Algorithm: A dual-wavelength PPG processing algorithm designed for 25 Hz sampling, performing beat-by-beat extraction and rolling window aggregation for $R$-SpO2 calibration.
• Motion-Aware Beat Selection: A strategy that utilizes accelerometer data to assign reliability weights to beats, improving robustness against micro-perturbations by prioritizing low-motion data points.
• Perfusion-Guided Calibration: A method that uses the perfusion index as a calibration-state indicator, allowing the $R$–SpO2 relationship to adapt to subject-specific variations in local perfusion conditions.

Experimental Results
The framework was evaluated on a private wearable dataset comprising 27 sessions from nine participants with diverse skin tones. The dataset included 100 Hz PPG signals (downsampled to 25 Hz for evaluation) and reference SpO2 values from a Masimo Rad-G device.

• Performance: Under a strict leave-one-subject-out (LOSO) evaluation, the proposed method achieved a Mean Absolute Error (MAE) of $2.305 \pm 1.113\%$ and a Root Mean Squared Error (RMSE) of $3.117 \pm 1.743\%$ at 25 Hz. This outperformed baseline methods including simple mean/median aggregation, Recursive Least Squares (RLS), and Normalized Least Mean Square (NLMS).
• Ablation Study: Results confirmed that both motion-aware beat selection and perfusion-guided calibration independently improved performance, with the combination yielding the best metrics across both 25 Hz and 100 Hz sampling rates.
• Low-Rate Efficiency: The method maintained performance at 25 Hz comparable to 100 Hz. Power consumption analysis indicated that the PPG sensor consumed approximately 9 mW at 25 Hz, a 40% reduction compared to the ~15 mW consumed at 100 Hz, supporting energy-efficient continuous monitoring.
• Motion Robustness: The combined approach demonstrated the lowest MAE across all motion magnitude levels, whereas baseline methods showed increased error as motion magnitude standard deviation increased.

Significance
The paper claims that this work presents a lightweight framework capable of improving the robustness of low-rate wrist SpO2 estimation without requiring high-complexity denoising models. By integrating motion-aware selection and perfusion-guided calibration, the system addresses the dual challenges of micro-perturbations and inter-subject perfusion variability. The ability to maintain accuracy at 25 Hz while significantly reducing power consumption makes the framework suitable for firmware-compatible, energy-efficient wearable implementations for continuous health monitoring.