OBJECTIVES: ML predictions of healthcare expenditures need to be fair across sensitive attributes such as gender, race and ethnicity, socioeconomic status, and geographic location to achieve health equity. The objective of this study is to quantify how fair the algorithm is in its predictions of healthcare expenditures for sensitive attributes.

METHODS: A retrospective cohort analysis with one-year baseline and one-year follow up period of older adults with incident colorectal cancer survivors diagnosed between 2010 and 2017 (N = 32,702) from SEER-Medicare database. Total expenditures(2018$) were derived from Medicare payments for inpatient, outpatient, prescription drugs, home health, and others. Sensitive attributes comprised sex, geographic location, poverty status, and race and ethnicity. eXtreme Gradient Boosting(XGBoost) and Shapley additive explanations with patient-, county-, and census tract- level features (N = 86) were used. Fairness was evaluated using differences in RMSE, R-square, MAE across subgroups and counterfactual using flip tests. Subgroup differences in these metrics between -0.1 and +0.1 were considered fair.

RESULTS: The sensitive attributes distributions were: 58% female, 9% non-metro, 31% poor, 8% NHA, 6% Hispanic, and 7% AA/PI/AN and 1% other race. The average healthcare expenditures was ($66753 + $53770). The incremental expenditures were: $-5,463(females), $-9,877(non-metro), $15,881(poor), $8,943(NHB), $8,143(Hispanic), $3,945(AA/PI/AN), $8,999(multiple race). For the overall model, the RMSE and R-Square for test dataset were 0.65 and 0.50. The top 4 leading predictors were: multimorbidity, cancer treatment (chemotherapy and surgery), and localized cancer stage. RMSE for NHB, poor, and males were higher compared to NHW, not poor, and females. Counterfactual fairness did not reveal any differences in fairness metrics.

CONCLUSIONS: Fairness metrics of XGBoost models suggested bias in predictions for NHB, males, and poor colorectal cancer survivors. Predictive ML models need to routinely report fairness metrics for sensitive attributes along with model performance and leading predictors.