Enhanced SpO₂ estimation using explainable machine learning and neck photoplethysmography.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Wearable Technology

Background:

• Reflectance-based photoplethysmogram (PPG) sensors offer versatile sites for blood oxygen saturation (SpO₂) monitoring.
• Accuracy limitations arise from diverse human characteristics and sensor settings, particularly at non-standard measurement locations.

Purpose of the Study:

• To develop an automated method for establishing subject inclusion-exclusion criteria for SpO₂ devices using reflectance-based PPG sensors.
• To enhance SpO₂ estimation accuracy at non-standard measurement sites by addressing variations in subjects and device settings.

Main Methods:

• Utilized unsupervised clustering, supervised regression, and SHAP (SHapley Additive exPlanations) for automated criteria construction.
• Developed a pillow-based wearable device for collecting reflectance PPG signals from neck arteries (brachiocephalic and carotid).
• Conducted experiments on 33 subjects with 80 different sensor settings per subject.

Main Results:

• The proposed approach effectively addressed human variability, device settings, and signal-to-SpO₂ mapping heterogeneity.
• Identified optimal device settings and applicable subject groups, defining inclusion-exclusion criteria.
• Achieved a 16% reduction in root mean squared error (RMSE) compared to empirical and plain estimation models.

Conclusions:

• The automated method successfully improves SpO₂ estimation accuracy at non-standard sites.
• The study demonstrates the utility of SHAP in explaining unsupervised learning clusters for device calibration.
• This approach enhances the reliability and applicability of reflectance-based PPG sensors for SpO₂ monitoring in diverse populations and conditions.