Enhanced SpO2 estimation using explainable machine learning and neck photoplethysmography

Abstract

Reflectance-based photoplethysmogram (PPG) sensors provide flexible options of measuring sites for blood oxygen saturation (SpO₂) measurement. But they are mostly limited by accuracy, especially when applied to different subjects, due to the diverse human characteristics (skin colors, hair density, etc.) and usage conditions of different sensor settings. This study addresses the estimation of SpO₂ at non-standard measuring sites employing reflectance-based sensors. It proposes an automated construction of subject inclusion-exclusion criteria for SpO₂ measuring devices, using a combination of unsupervised clustering, supervised regression, and model explanations. This is perhaps among the first adaptation of SHAP to explain the clusters gleaned from unsupervised learning methods. As a wellness application case study, we developed a pillow-based wearable device to collect reflectance PPGs from both the brachiocephalic and carotid arteries around the neck. The experiment was conducted on 33 subjects, each under totally 80 different sensor settings. The proposed approach addressed the variations of humans and devices, as well as the heterogeneous mapping between signals and SpO₂ values. It identified effective device settings and characteristics of their applicable subject groups (i.e., subject inclusion-exclusion criteria). Overall, it reduced the root mean squared error (RMSE) by 16%, compared to an empirical formula and a plain SpO₂ estimation model.

Keywords: Explainable machine learning; Neck reflectance photoplethysmogram (PPG); SpO(2) estimation; Subject heterogeneity; Subject inclusion-exclusion criteria.