Evaluating fairness and mitigating bias in models predicting financial toxicity among patients with genitourinary cancers.

Abstract

1565 Background: Financial toxicity, the economic burden patients face from healthcare expenses, is a growing concern in cancer care. Recognizing the high costs of diagnosis and treatment of genitourinary (GU) cancers, this study aims to (1) comprehensively characterize the socioeconomic, demographic, and care-related factors associated with financial toxicity in patients with GU cancers, and (2) evaluate bias in the predictive model developed using these patient factors. Methods: The 2019–2022 Medical Expenditure Panel Survey (MEPS) data was used to identify patients with GU cancers. MEPS captures utilization, frequency, cost, and payment sources of U.S. health services alongside health insurance coverage characteristics and accessibility in the workforce. Financial toxicity was defined as patient-reported difficulties paying medical bills, high out-of-pocket expenses ( > 10% of total income), and high self-pay ratios ( > 20% of total healthcare expenditure). Predictive modeling was performed using logistic regression using age, sex, race/ethnicity, income, insurance status, and expenditure-related predictors. To address potential algorithmic bias, Fairlearn's ThresholdOptimizer, a postprocessing algorithm, was applied to this predictive model, adjusting predictions to ensure equalized odds across racial groups. Performance metrics, including accuracy, precision, and recall, were evaluated overall and by racial group. Results: Overall, we identified 1131 patients with GU cancers (weighted n = 11,723,024) in the MEPS data; median age 72 yrs; sex 93.4% male; 71.6% White, 18.3% Black, 6.5% Hispanic, and 3.5% Other. 22.2% of patients reported financial toxicity with a median [Q1, Q3] total healthcare expenditure of $2,645.0 [$898.5, $5328.0] vs. $503.5 [$171.0, $1286.8]. Logistic regression achieved an overall accuracy of 95%, with a precision of 97% and recall of 77% for financial toxicity cases. Fairness metrics of the unadjusted predictions revealed bias to specific communities with lower recall for Black (46.2%) and Other Races (33.3%) compared to Hispanic (75.0%) and White (90.4%) patients. After threshold optimization, recall improved to 61.5% for Black and 50% for Other Races, while Hispanic (84.6%) and White (100%) patients maintained high performance. However, disparities persisted, as evidenced by an equalized odds difference of 0.21. Conclusions: This study underscores the critical need for responsible development of predictive models impacting cancer care. Our findings show that a bias-correcting postprocessing algorithm can be an essential tool since it can be applied to existing models without requiring retraining; however, these algorithms do not represent a definitive solution since this model’s underlying bias persists, highlighting the need to ensure models learn from fair data sets that are representative of the US population.