OBJECTIVES: To develop a machine learning (ML) model to predict 5-year Cancer-Specific Survival (CSS) and quantify drivers of racial/ethnic disparities among patients with hormone receptor-positive breast cancer (HR+ BC).

METHODS: This retrospective cohort study used the Surveillance, Epidemiology, and End Results Program (SEER) cancer registry data from female patients who were diagnosed with HR+ BC between 2010-2016, and county-level SocioEconomic Status Index (SES) data from the CDC. We predicted a 5-year CSS utilizing an XGBoost ML model using 12 variables related to cancer characteristics, surgery, and demographic factors and measured its performance using concordance index (C-index). We used SHapley Additive exPlanations (SHAP) values to explain the impact of specific variables on hazard predictions, including across racial/ethnic subgroups.

RESULTS: Among 259,305 HR+ BC patients, 69% were Non-Hispanic White (NHW), 11% Hispanic, 10% non-Hispanic Black (NHB) and 10% were Other Non-Hispanic.The ML model predicted 5-year CSS well with a C-index of 0.89±0.003. Of what could be explained of CSS hazards, cancer staging explained 50.7% of the variation, followed by surgery type (15.1%), age at diagnosis (10.7%), insurance type (6.5%), marital status (4.9%), SES (4.5%), and race/ethnicity (3.2%). Compared to NHW patients, mortality risk was similar in Hispanic patients but higher in NHB patients with a hazard ratio of 1.55 (95%CI: [1.42, 1.69]). The latter racial disparity was largely mediated by differences in cancer stage at diagnosis (36.7%), surgery type (10.8%), marital status (6.4%), SVI (6.2%), and insurance type (4.9%).

CONCLUSIONS: A ML approach was able to accurately predict 5-year CSS. Cancer staging was estimated to be the biggest driver of disparity between NHB and NHW. This suggests efforts to detect cancer earlier among NHB women may help to narrow the gap in survival.