ML4 Association of Incident Cancer with Low-Value Care among Elderly Medicare Beneficiaries Using Machine Learning


      In the United States (US), 25% of healthcare spending is considered wasteful because it is spent reimbursing low-value care. Low-value care is the utilization of healthcare services, medical tests, and procedures that have unclear or no clinical benefit to patients, but still exposes them to risk. This study aims to evaluate the association of incident breast, prostate, colorectal and Non-Hodgkin’s cancer to low-value non-cancer care among older US adults enrolled in Medicare using machine learning methods.


      We used a retrospective cohort study design with 12-month baseline and follow-up periods. We identified two cohorts of cancer and non-cancer patients. We identified relevant low-value services using ICD9/ICD10 and CPT/HCPCS codes. XGboost models were used to identify the leading predictors of low-value care and partial dependence plots to examine the association of the different cancer types to low-value care.


      The combined study cohorts included 529,452 individuals. Overall, the prevalence of low-value care was 24.3%. Rates of low-value care differed significantly by cancer type; the highest rates were observed in Non-Hodgkin’s lymphoma (34%) followed by colorectal cancer (29% ) while the lowest rates were among patients diagnosed with prostate cancer (22%). The association of cancer to low-value care varied by cancer type; both colorectal cancer and NHL were positively associated with low-value care, but breast and prostate cancers were negatively associated with low-value care.


      One in four older fee-for-service Medicare beneficiaries received low-value care. The leading patient-level predictors of low-value care were fragmentation of care, the number of chronic conditions, and age. Community-level predictors like market characteristics, healthcare utilization, and social determinants of health were also found to be important predictors of low-value care, suggesting that a multipronged approach that targets patient and system-level factors are needed to reduce the risk of low-value care among older adults.