Background: Oxygen saturation (SpO2) is often overestimated in darker skinned individuals (1–3), but the ideal method to quantify this phenomenon and measure its impact on clinical outcomes is unclear. Studying the distributions of SpO2 may not be the appropriate method because these distributions are influenced by clinicians’ real time efforts to maintain SpO2 in a particular range (typically 90-94%) by adjusting patients’ supplemental oxygen levels. We hypothesized that the ratio of the partial pressure of arterial oxygen (PaO2) to fraction of inspired oxygen (FiO2) – the P/F ratio – would be a better way to study the impact of falsely elevated oximetry readings on clinical outcomes. We developed a simple algorithm to non-invasively estimate the P/F ratio without arterial blood gases (estimated P/F ratio, or ePFR).
Methods: In a retrospective cohort of COVID-19 hospital encounters (n =1100; between 03/2020 and 02/2021 at one academic tertiary hospital), we characterized racial disparities using SpO2 and ePFR. We used a model of the oxygen dissociation curve (4) to estimate PaO2 from SpO2 (Fig. 1A). We used nurse-recorded supplemental oxygen settings to determine FiO2 (Fig. 1B). We computed empirical cumulative distribution functions (ECDFs) for SpO2 and ePFR, and quantified racial differences in the distributions using the Kolmogorov-Smirnov (K-S) distance. We used adjusted odds ratio (aOR) and area under receiver operator characteristic curve (AUROC) from multivariable logistic regression to characterize the influence of race on the relationship between hypoxemia severity estimators (SpO2 and ePFR) and the primary outcome (death or transfer to intensive care unit within 24 hours).
Results: The primary outcome occurred in 17% of our 1100 patients and 1.85% of our 399,797 rows of data (every 15 minutes). The median and inter-quartile range for SpO2 and ePFR were 96% (94-98%) and 321 (241-395) respectively. A wider racial disparity was revealed by the ECDF of the ePFR (K-S distance = 0.14, p < 0.01) than that of SpO2 (K-S distance = 0.07, p < 0.01) (Fig. 2A & 2B). The influence of race was better demonstrated by ePFR-based models than SpO2-based models of clinical deterioration (Fig. 2C & 2D). When race (Black vs non-Black) and SpO2 were added as predictors to a baseline risk model (age, sex, and Charlson comorbidity index), the improvement in model discrimination was not clinically significant (AUROC of 0.64 vs 0.60 for baseline model, p < 0.01) and influence of race was not detected (p = 0.14). In contrast, when race and ePFR were added as predictors to the same baseline model, the improvement in model discrimination was clinically significant (AUROC of 0.75 vs 0.60 for baseline model, p < 0.01) and influence of race was obvious (aOR of 0.5 for non-Black patients, p < 0.01).
Conclusions: Racial bias in pulse oximetry has been known for decades. Yet, the problem has been poorly studied and remains uncorrected. This study validates the ePFR as a tool to better demonstrate the effects of racially biased pulse oximetry readings. The proposed algorithm lends itself to convenient implementation in large data sets and/or electronic medical records. It can thereby allow health systems to create more equitable predictive models for their patients. Further, this tool can empower healthcare and/or political leadership to confront pulse oximetry manufacturers with evidence of racial bias and create the market forces and/or regulatory climate needed to bring an end to this important source of structural inequity.
