2A-05 - Evaluating the Impact of Varying Bias Thresholds in Pulse Oximeter Performance on Racial and Ethnic Disparities: A Simulation Study in COVID-19 Patients

Background: Pulse oximeters often overestimate oxygen levels in individuals with darker skin, potentially driving racial and ethnic disparities in clinical care. It’s unclear how different thresholds of “acceptable” measurement bias impact disparities. Determining clinically acceptable bias levels without worsening disparities is key to developing regulatory guidelines that promote equitable pulse oximetry performance.

Objectives: To evaluate racial and ethnic bias in SpO₂ measurement and quantify the impact of different thresholds for “acceptable” bias on disparities in outcomes.

Methods: We conducted a target trial emulation using electronic health record of COVID-19 patients treated at Johns Hopkins Health System from March 2020 to June 2024. Race and ethnicity were self-reported and categorized based on 2024 Office of Management and Budget standards, with the two groups of focus being non-Hispanic Black (NHB) and White (NHW). The exposure was SpO₂ measurement error, defined as the difference between SpO₂ and SaO₂. Hypothetical interventions were defined by different thresholds of racial and ethnic bias in SpO₂ measurement, with each scenario assigned a distribution of SpO₂ error. The outcomes were supplemental oxygen (O₂) and dexamethasone use in 24 hours and hospital readmission in 30 days. We used G-computation to estimate average potential outcomes under different bias scenarios, standardizing results using the distribution of allowable covariates (e.g., measures of clinical indications) in NHB patients.

Results: A total of 15,122 SpO₂-SaO₂ pairs (7,350 NHB; 7,772 NHW) showed a greater measurement error among NHB (mean: 1.99% vs. 0.56%, P < 0.001) and a higher probability of occult hypoxemia (4.80% vs. 2.34%, P < 0.001). NHB and NHW patients had similar probability of receiving supplemental O₂ (98%), with no significant differences across intervention scenarios. NHW patients had a higher probability of receiving dexamethasone than NHB (28.19% vs. 20.05%), with a disparity of -8.14% (95% CI: -15.60%, -1.12%), which was decreased after adjusting for allowables. Hospital readmission rates were higher in NHB (14.10% vs. 8.73%), with a persistent disparity (~5%) across intervention scenarios.
The study cannot generalize beyond COVID-19 patients within a single health system and may have residual bias from unmeasured clinical need and provider reliance on SaO₂ for clinical decision-making. Future work should evaluate impacts on treatment timing in broader populations.

Conclusions: Varying our specified thresholds of racial and ethnic bias in SpO₂ measurement had no impact on disparities in treatment outcomes, while disparities in hospital readmission persisted.