Overcoming inequities in Pulse oximetry Through clinical InformatiCs (OPTIC)

Duke University, Durham NC

Investigators

Abstract

Background/Significance: Pulse oximetry is the most common method to monitor arterial oxygen saturation in patients in modern medicine. This technology reflects light off oxygenated blood to approximate the more accurate arterial blood gas (ABG) tests. Recent findings demonstrate that pulse oximeters are more inaccurate in patients with darker skin tones, likely due to interference by skin tone. Minority patients consequently experience disproportionately more hidden hypoxemia (HH), which occurs when pulse oximetry overestimates oxygen levels as within the normal range (≥ 88%) but demonstrate hypoxemia by the more accurate ABGs. However, ABGs are painful to administer and only provide episodic data, so they are used less frequently than pulse oximetry. Furthermore, minority patients are less likely to receive ABG tests, making it even more likely that the problem is underestimated. Innovation: Replacing our pulse oximeters will require significant investment and time. However, a two- stage alternative solution is possible now, leveraging existing pulse oximetry hardware. HH causes patients to have internal organ malfunction, which is reflected in abnormal vitals and laboratory values. We will develop solutions in this multi-health system study (Duke Health, Emory Healthcare) spanning 9 hospitals. In Aim 1, we use routinely-collected electronic health record (EHR) data to develop a machine learning (ML) model to predict which patients are at high risk for HH. Furthermore, we use informative sampling to identify which patients may best benefit from an ABG. Some of these informative patients may benefit from skin tone measurement. In Aim 2a, we measure skin tone using multiple methods - administered visual scales (Fitzpatrick, Monk) and color measurement devices (colorimeters, spectrophotometers, mobile phone cameras) - across patients in hospitals to predict the degree of pulse oximetry bias in critical illness. Using informative sampling in Aim 2b, we build on Aim 2a patients to iteratively identify and test which factors best identify skin tones that will best improve our ML models. Finally, we prospectively perform silent validation for Aims 1 and 2 models in Year 5. Next steps: The results of this proposal will build silently validated, implementation-ready ML models that can mitigate the impact of HH using current pulse oximetry hardware and EHR data. Furthermore, by measuring skin tone, it can mitigate the impact of skin tone on pulse oximetry bias through new ML models when combined with EHR data. Finally, this project will conduct silent validation to ensure temporal validity and understand the impact within current workflows. Completing this project will form the foundation for two future R01 proposals to conduct cluster randomized trials to test the implementation of EHR-only (Aim 1) and skin-tone augmented EHR ML models (Aim 2) to mitigate the effects of pulse oximetry inequity in current clinical practice.