Machine Learning Risk Stratification for Older Breast Cancer Survivors: Clinical Care Implications.

OBJECTIVE: To develop and validate a clinical risk prediction algorithm to identify breast cancer survivors at high risk for adverse outcomes. STUDY SETTING AND DESIGN: Our national retrospective analysis used cross-validated random forest machine learning models to separately predict the risk of all-cause death, cancer-specific death, claims-derived risk of recurrence, and other adverse health outcomes within 3 and 5 years following treatment completion. DATA SOURCES AND ANALYTIC SAMPLE: Our study used the Surveillance and Epidemiology End Results (SEER) registry-Consumer Assessment of Healthcare Providers and Systems (CAHPS) survey (SEER-CAHPS) linked data for survivors diagnosed between 2003 and 2011, with follow-up claims data to 2017. PRINCIPAL FINDINGS: Within the 3-year follow-up period, 372/4516 survivors (mean age 75.1; 81.7% white) in the primary cohort (8.2%) died, 111 from cancer (2.5%), 665 (14.7%) experienced cancer recurrence, and 488 (10.8%) were hospitalized for adverse health outcomes. The algorithm's prediction resulted in 91.9% out-of-sample accuracy (the percent of observations classified correctly) and a 37.6% Cohen's Kappa (i.e., improvement over an uninformed model). Out-of-sample accuracy was 97.5% (44% improvement) for predicting cancer-specific death, 85% (26% improvement) for recurrence, and 89% (28% improvement) for other adverse health outcomes. Important predictors across outcomes included geographic region, age, frailty, comorbidity, time since diagnosis, and out-of-pocket cost responsibility. CONCLUSIONS: Machine learning models accurately predicted relevant adverse survivorship outcomes, driven primarily by non-cancer specific factors. Breast cancer survivors at high risk for adverse outcomes may benefit from more intensive care, whereas those at low risk may be more appropriately managed by primary care.

Duke Scholars

Author Michaela Ann Dinan Population Health Sciences