Pulse oximetry is an indispensable, non-invasive tool for monitoring functional arterial oxygen saturation (SpO2) across diverse clinical settings. Its accuracy, however, remains limited by discrepancies between SpO2 and arterial blood gas-derived oxygen saturation (SaO2), the clinical gold standard. These inaccuracies arise from both technical factors (e.g., motion artifacts, device variability) and physiological influences (e.g., skin pigmentation, peripheral perfusion, underlying health conditions). Moreover, SpO2 measurements vary across anatomical sites, and no consensus exists on the optimal sensor location, as performance depends heavily on contextual and patient-specific factors.
The thesis addresses the research question: Can predictive models, leveraging physiological and contextual features together with multi-site pulse oximetry data, enhance SpO2 estimation accuracy compared to conventional single-site measurements? To investigate, predictive models were developed for eight sensors using data from two desaturation studies (SaO2 70-100%). Parsimonious regression models were designed to predict SpO2 bias (SpO2-SaO2) with minimal informative feature sets. Corrected Spo2 values derived from these models were further integrated through three multi-sensor fusion strategies.
Results demonstrated significant improvements in bias prediction for 3 of the 8 sensors, with 7 showing reductions in ARMS (accuracy root mean square), ranging from 4.43% to 37.02% relative to baseline. The quadratic weighting fusion method, which weighted corrected SpO2 values inversely to the square of their predicted bias, achieved statistically significant improvements in 21 of 28 sensor pairs, while consistently reducing ARMS across all combinations. Importantly, these models also reduced differential bias, mitigating the systematic overestimation of SpO2 in individuals with darker skin tones.
This work demonstrated that sensor-specific bias correction models and context-aware multi-site fusion can substantially improve both accuracy and fairness in SpO2 monitoring. While the study is limited by its reliance on healthy volunteers, a small dataset, and a restricted set of sensors, the framework provides a promising foundation. Future work should focus on clinical validation in diverse patient populations. With further development, these algorithms could be embedded into oximeter devices, enabling real-time, patient-specific bias correction and paving the way toward more reliable and equitable pulse oximetry.