ALGORITHMIC FAIRNESS IN MACHINE LEARNING PREDICTION FOR LONG-TERM ADJUVANT ENDOCRINE THERAPY ADHERENCE AMONG BREAST CANCER SURVIVORS
Author(s)
Yeijin Kim, MS, PharmD, Li-Wei Wu, MS, Nishamathi Kumaraswamy, PhD, Chanhyun Park, MEd, RPh, PhD;
The University of Texas at Austin, Austin, TX, USA
The University of Texas at Austin, Austin, TX, USA
OBJECTIVES: Adherence to adjuvant endocrine therapy (AET) remains inequitable among older women with hormone receptor-positive breast cancer (HR+ BC). We evaluated the prediction performance and algorithmic bias of a machine learning (ML) model for AET adherence across the domains of the Minority Health Social Vulnerability Index (MHSVI).
METHODS: This U.S. population-based cohort study included women aged ≥ 65 years who were diagnosed with HR+ BC and initiated AET between 2013 and 2017, using 2007-2022 SEER-Medicare data. Women were stratified by four MHSVI quartiles. The outcome was 5-year binary adherence (proportion of days covered [PDC] ≥ 0.80). We developed an XGBoost model using LASSO-selected features, with MHSVI scores (overall and six themes) as primary predictors. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC). Algorithmic fairness was assessed using equal opportunity difference (EOD), demographic parity difference (DPD), false negative rate (FNR), and false positive rate (FPR), with the lowest-vulnerability group as the reference.
RESULTS: We identified 38,320 women (mean age = 75.0 years, SD = 5.8). The XGBoost model achieved a ROC-AUC of 0.63. Significant fairness gaps were observed: the EOD between the most and the least vulnerable groups was -0.15, and the DPD was -0.12. Among the six MHSVI themes, higher vulnerability was significantly associated with increased FNR in the Socioeconomic theme (0.30 in the lowest vs. 0.43 in the highest vulnerability group; p < 0.05) and Minority Status and Language theme (0.19 vs. 0.35; p < 0.05). The Medical Vulnerability theme showed a similar trend in FPR (0.46 to 0.58; p < 0.05).
CONCLUSIONS: Traditional ML models may disproportionately misclassify patients at risk of non-adherence, particularly among those with high social vulnerability. These disparities in prediction errors highlight the need for fairness-adjusted ML algorithms to ensure prediction equity and more equitable resource allocation for AET support interventions.
METHODS: This U.S. population-based cohort study included women aged ≥ 65 years who were diagnosed with HR+ BC and initiated AET between 2013 and 2017, using 2007-2022 SEER-Medicare data. Women were stratified by four MHSVI quartiles. The outcome was 5-year binary adherence (proportion of days covered [PDC] ≥ 0.80). We developed an XGBoost model using LASSO-selected features, with MHSVI scores (overall and six themes) as primary predictors. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC). Algorithmic fairness was assessed using equal opportunity difference (EOD), demographic parity difference (DPD), false negative rate (FNR), and false positive rate (FPR), with the lowest-vulnerability group as the reference.
RESULTS: We identified 38,320 women (mean age = 75.0 years, SD = 5.8). The XGBoost model achieved a ROC-AUC of 0.63. Significant fairness gaps were observed: the EOD between the most and the least vulnerable groups was -0.15, and the DPD was -0.12. Among the six MHSVI themes, higher vulnerability was significantly associated with increased FNR in the Socioeconomic theme (0.30 in the lowest vs. 0.43 in the highest vulnerability group; p < 0.05) and Minority Status and Language theme (0.19 vs. 0.35; p < 0.05). The Medical Vulnerability theme showed a similar trend in FPR (0.46 to 0.58; p < 0.05).
CONCLUSIONS: Traditional ML models may disproportionately misclassify patients at risk of non-adherence, particularly among those with high social vulnerability. These disparities in prediction errors highlight the need for fairness-adjusted ML algorithms to ensure prediction equity and more equitable resource allocation for AET support interventions.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
PCR185
Topic
Patient-Centered Research
Topic Subcategory
Adherence, Persistence, & Compliance
Disease
SDC: Oncology