Abstract Global budgets and risk-based healthcare payments depend on accurate risk adjustment. The CMS Hierarchical Condition Category (CMS–HCC) model links demographic and diagnostic information to predicted spending to support these arrangements. However, if diagnostic coding intensities vary across organizations, predicted payments may deviate from underlying morbidity. This study introduces an after-the-fact, supplementary consistency check that uses low-discretionary hospital admissions as an independent reference for detecting organization-level variation in risk-score coding intensity. We conducted a computational simulation and evaluation study using historical Medicare claims drawn from a standard sample. Study-specific CMS–HCC v22 spending predictions based on 2018 indicators were used to define predicted-spending deciles. Risk-score coding variation was simulated by probabilistically switching selected base-year CMS–HCC indicators from 0 to 1 within nine categories commonly identified as vulnerable to upcoding, at 10% and 20% intensification levels. A set of 45 low-discretionary DRG clusters was identified from 2019 inpatient claims and used in a decile-level DRG regression to generate supplementary spending predictions and corresponding error bars defined as ± 1.96 standard deviations (SD) of the DRG-based spending predictions. Both the DRG and CMS–HCC-based models predict total annual Medicare care spending in 2019, and all data pertain to members enrolled in traditional Medicare in both years. In the absence of simulated coding variation, validation cohort DRG-based predictions closely tracked CMS–HCC predictions and remained within the DRG-based prediction error bars. Under simulated coding intensification, CMS–HCC-based decile predictions rose above both actual spending and the DRG-based consistency estimates, exceeding + 1.96 SD at 10% intensification in all but the highest decile and at 20% intensification in all deciles. DRG-based decile prediction SDs ranged from 3.5% of the mean in the highest-spending decile to approximately 11% in the lowest-spending decile. Low-discretionary DRG clusters can provide a practical, group-level reference signal for detecting variation in CMS–HCC coding intensity. The method is intended as a supplementary consistency check rather than a stand-alone risk model and is naturally aligned with organization-level evaluations.
Crowley et al. (Fri,) studied this question.