Racial Disparities in Pulse Oximetry Cannot Be Fixed With Race-Based Correction

Pulse oximeters are vital for life-saving care, particularly during the COVID-19 pandemic, influencing decisions regarding emergency care and treatment eligibility. However, a significant racial bias in pulse oximetry has persisted for over 30 years, with studies consistently showing higher SpO2 readings for Black patients, even at dangerously low oxygen levels. This bias leads to occult hypoxemia (missed detection of low oxygenation) being more frequent in Black patients compared to white patients, resulting in unequal access to care. While the term "bias" can refer to both statistical bias (the difference between the expected pulse oximetry measurement and the actual arterial oxygenation saturation) and differential assessment by group, this research focuses on the statistical bias and its racial disparities. The paper addresses the limitations of race-based correction factors which have historically been problematic in healthcare, and aims to demonstrate that no single race-based correction will achieve equal performance in hypoxemia detection across races.

Existing literature extensively documents racial bias in pulse oximetry, highlighting the higher average SpO2 measurements in Black patients compared to white patients, particularly at lower oxygen saturation levels. Studies report significantly increased odds ratios for occult hypoxemia in Black patients. The consequences include delayed or inadequate care for Black patients experiencing equally dangerous low blood oxygenation levels, contributing to health disparities and potentially increased mortality. The issue extends beyond statistical bias to encompass racial discrimination in assessment, a phenomenon described as "health care gaslighting."

This retrospective analysis utilizes the eICU Collaborative Research Database (eICU-CRD), a publicly available dataset containing anonymized records from over 139,000 unique patients. The researchers extracted 218,000 pairs of SpO2 and SaO2 measurements, ensuring temporal proximity (within 10 minutes) and excluding patients with "other/unknown" race classifications. The five racial/ethnic categories included were white, African American/Black, Hispanic, Asian, and Native American. The study analyzed the statistical bias (SpO2 - SaO2), comparing the mean and median differences across racial groups and SpO2 ranges (low, medium, high). Further, it investigated the variance of the SpO2 error, utilizing Welch's t-tests, F-tests, and tests of equal proportions to assess statistical significance. Finally, the study examined the performance of a hypoxemia detector using different thresholds (γ) to evaluate the impact of racial disparities on hypoxemia detection, comparing probabilities of detection and false alarms across races.

The study confirmed a statistically significant higher bias in SpO2 measurements for Black and Asian patients compared to white patients. The bias varied as a function of SpO2 levels, with racial disparities widening at lower oxygen saturation. Importantly, the standard deviation of the SpO2 error was consistently higher for Black and Asian patients, indicating greater variability and imprecision in SpO2 measurements. Probability mass functions (pmfs) of SpO2 error revealed heavier tails for Black patients, signifying a higher likelihood of large errors. Notably, Black patients had a 67% higher chance of a large SpO2 error (>10) compared to white patients. Analysis of hypoxemia detection performance showed that no single threshold could achieve equal performance for Black and white patients. Race-based corrections, using different thresholds for different races, also failed to eliminate disparities in detection performance. While a potential race-based correction might be feasible for Asian patients, this does not address the more significant and pervasive issue for Black patients. The study underscores that the higher variance and heavier tails in the distribution of SpO2 errors for Black patients significantly impact hypoxemia detection, and this cannot be corrected by simply adjusting thresholds.

The study's findings challenge the prevailing notion that race-based corrections can solve the racial disparities in pulse oximetry. The results demonstrate that simple adjustments to thresholds are insufficient to address the fundamentally flawed nature of the device itself. The observed differences in variance and distribution shape, especially the heavier tails observed for Black patients, highlight a deeper problem of imprecision in SpO2 readings for this population. Melanin concentration in skin, affecting light attenuation during pulse oximetry, likely contributes to the observed variance differences. Furthermore, the social construction of race in the US likely plays a role, potentially leading to increased variability in melanin levels within the Black population. The study emphasizes the urgent need for redesigning pulse oximeters to eliminate racial bias. Addressing both the mean and variance of SpO2 errors is critical for developing equitable monitoring devices.

This study demonstrates that race-based corrections cannot adequately address the racial disparities in pulse oximetry. The inherent inaccuracies and imprecision of SpO2 measurements in patients of color, especially Black patients, necessitates a fundamental redesign of the device. Future research should include larger and more diverse datasets that include skin color as a variable, enabling a more granular understanding of the relationship between melanin, light attenuation, and SpO2 variance. Regulatory changes are also required to ensure that device testing incorporates representative populations of color, enhancing the accuracy and equity of pulse oximetry.

The study's reliance on existing data introduces limitations. The eICU-CRD has a disproportionately high number of white patients, hindering a more robust analysis of racial disparities in smaller groups (Asian and Native American). Further research with larger, more diverse datasets, including explicit measurement of skin color, is necessary to draw more definitive conclusions and explore potential confounding factors.