Resampling for Mitigating Bias in Predictive Model for Substance Use Disorder Treatment Completion

While predictive models in healthcare have great potential for discovering knowledge and supporting decision making for clinicians, biases in these models can cause great harm, including exacerbating existing health disparities or introducing new ones. Completion of substance use disorder (SUD) treatment is a significant predictor of future outcomes for patients. The ability to predict beforehand which patients are at higher risk of not completing treatment would offer clinicians an opportunity to intervene early in the process and potentially result in better outcomes. The objective of this study was to investigate biases in a model for SUD treatment completion with respect to the sensitive attributes of race and sex, and then enhance model fairness using data pre-processing techniques, specifically resampling. Models trained on the data before resampling exhibited significant racial bias and moderate sex bias. Sequential and simultaneous undersampling approaches effectively improved model fairness without compromising performance as measured by AUROC and sensitivity, and are thus appropriate bias mitigation techniques for this task.