Evaluating Equity in XGBoost Predictions of High Healthcare Expenditures for Older Women with Osteoarthritis in the United States

Purpose:

Osteoarthritis (OA) is a highly prevalent and debilitating condition among older adults. Studies suggest that women are more prone to develop symptomatic disease than men. OA is associated with high direct healthcare costs, attributable to its complex disease management. Furthermore, a small segment of this population may incur very high costs. Identifying these high-cost users is important for allocation of resources, cost containment, quality improvement, and population health management.

However, current research in the prediction of high-cost users in OA using machine learning (ML) models is limited. Furthermore, ML model predictions of high-cost users must be equitable across sensitive attributes such as race and ethnicity and socio-economic status.

This study investigated the leading predictors of high-cost users among older women with OA utilizing ML methods and the fairness of the ML algorithm in its predictions across subgroups of race and ethnicity, poverty, and education.

Methods:

A cross-sectional study was conducted using data from older women (age>65 years) with OA from the 2021 Medical Expenditure Panel Survey, a nationally representative survey of the non-institutionalized civilian households in the US. High-cost users were identified as having higher than the 90th percentile (>$39,388) in total healthcare expenditures. Key predictors were identified using interpretable ML model eXtreme Gradient Boosting (XGBoost) Classification and SHapley Additive exPlanations (SHAP). Overall model fit was evaluated with AUC, recall, and precision. Fairness was measured with demographic parity (equalization of odds, disparate impact, and equal opportunity) across racial and ethnic groups (Non-Hispanic White (NHW), Non-Hispanic Black (NHB), Hispanic ethnicity), education (no college and college) and poverty status (low income and high income). Counterfactual fairness was evaluated to ensure consistency in high-cost predictions between actual scenarios and counterfactual situations where individuals belong to different groups.

Results:

A higher percentage of Hispanic (12.2%) and NHB (14.4%) were high-cost users compared to NHW (9.0%). A higher percentage of older women without college education (10.7%) and with low income (11.2%) compared to those with college education (2.5%) and high income (5.2%) were high-cost users. The overall model fit was acceptable with AUC 0.81, recall 0.62, and precision 0.91. Multimorbidity, high school education level, and anxiety were the top 3 predictors of high-cost users. Prediction was lower among older women without college education (AUC = 0.80) and low income compared (AUC = 0.77) compared to overall prediction (AUC = 0.81). Demographic parity revealed little to no differences across racial and ethnic, education, and income groups.

Conclusion:

The fairness metrics indicated no bias in the predictions, likely attributable to the nationally representative nature of the survey sample and its large size. These findings need to be confirmed with other data that contain diverse populations. Leading predictors indicated that effective management of multimorbidity may reduce the risk of high-cost use in older women with OA.