Addressing Racial Disparities in Pulse Oximetry: A Machine Learning Perspective

Pulse oximetry is a widely used non-invasive method for estimating arterial oxygen saturation (SaO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>). However, it has been shown to exhibit racial biases in its blood oxygen saturation (SpO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>) measurements, primarily due to differences in skin color, which can lead to misdiagnosis and improper treatment. This study utilizes the publicly available BOLD dataset on PhysioNet to investigate these biases. It employs machine learning models, including Random Forest, Gradient Boosting, and XGBoost regressors, to predict SpO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> from SpO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> readings, incorporating interpersonal parameters such as race, with an initial focus on the most distinguishable groups by skin color: Black and White. The developed models significantly improve prediction accuracy, reducing the Mean Squared Error from 30.76 (baseline) to 4.72. SHAP analysis underscores the critical role of race as a surrogate for skin tone in enhancing predictions. These findings emphasize the need for equitable signal processing techniques for medical devices to decrease the bias and improve health outcomes across diverse populations.