From data to dollars: Evaluating the predictive power of enhancing oncology model risk adjusters.

Abstract

e13894 Background: The Enhancing Oncology Model (EOM) is an oncology-specific, total cost of care, value-based payment model. Performance in the model is assessed by comparing actual expenditures against predicted prices, with the goal of spending fewer dollars than predicted. Predicted prices, derived using cancer-specific prediction models, regress expenditures on a variety of covariates (risk adjusters) that have been determined to be associated with episode expenditures. These covariates include age, sex, dual eligibility, Medicare Part D enrollment, low-income subsidy, comorbidities (defined by Hierarchical Condition Categories), receipt of select cancer-directed surgeries, bone marrow transplants, radiation therapy, institutional status, clinical trial participation, prior chemotherapy use, and episode length. This study evaluates the predictive capability of these price prediction models. Methods: Utilizing deidentified national baseline episode data from the Center for Medicare & Medicaid Innovation, we performed regression analyses for each cancer-type using covariates defined in the EOM Payment Methodology. We then assessed the R-squared values, which indicate the proportion of variance in the dependent variable explained by the independent variables (a larger R-squared value indicating that the chosen covariates explain the expenditure variations better), for each linear model. Results: The analysis regressed 1.5 million episodes between July 2016 and June 2020 using 7 independent cancer-type specific linear regression models. For each cancer type, the intercept is large, covariates are statistically significant, F-statistic is high, but the regression model adjusted r-squared value is < 0.2, indicating variations that are unexplained by the chosen covariates. Conclusions: The findings suggest that while EOM's price prediction framework shows promise, it does not account for significant factors influencing episode expenditures. Models incorporating additional or different statistically significant risk adjusters may better explain the variation in expenditures and improve predictive accuracy. We recommend exploring clinical factors (such as disease stage, tumor characteristics, biomolecular markers, genomics, disease and performance status, treatment type and adherence, prognostic indicators, and comorbidities) or psycho-socio-economic covariates (such as the Area Deprivation Index or Social Vulnerability Index). Regression model outputs. Cancer Type Episode Count Intercept Covariate Count Adjusted R-squared Prob(F-Statistic) Breast Cancer 338,495 36,570 30 0.072 <0.001 Lung Cancer 343,611 58,660 27 0.014 Colorectal Cancer 181,480 27,480 31 0.083 Prostate Cancer 135,879 37,440 26 0.054 Lymphoma 194,612 56,120 43 0.208 Multiple Myeloma 217,855 58,440 33 0.111 Chronic Leukemia 121,488 47,450 33 0.085