Performance of comorbidity adjustment measures to predict healthcare utilization and expenditures for patients with diabetes using a large administrative database

Objective: The objective of this study was to compare the use of different comorbidity measures to predict future healthcare utilization and expenditures for diabetic patients. Methods: This was a retrospective study that included 8,704 diabetic patients enrolled continuously for three years in the Department of Defense TRICARE program. Administrative claims data were used to calculate six comorbidity measures: number of distinct medications, index-year healthcare expenditures, two versions of the Charlson Comorbidity Index (CCI), and two versions of the Chronic Disease Score (CDS). Linear regression models were used to estimate three health outcomes for one- and two-year post-index periods: healthcare expenditures (COST), number of hospitalizations (HOS), and number of emergency department visits (ED). Logistic regression models were used to estimate binary outcomes (above or below the 90th percentile of COST; [greater than or equal to] 1 HOS or none; [greater than or equal to] 1 ED or none). Comparisons were based on adjusted R², areas under the receiver-operator-curve (c statistics), and the Hosmer-Lemeshow goodness-of-fit tests. Results: The study population had a mean age of 51.0 years (SD = 10.5), and 46.3 percent were male. After adjusting for age and sex, the updated CCI was the best predictor of one-year and two-year HOS (adjusted R² = 8.1%, 9.3%), the number of distinct medications was superior in predicting one-year and two-year ED (adjusted R² = 9.9%, 12.4%), and the index-year healthcare expenditures explained the most variance in one-year and two-year COST (adjusted R² = 35.6%, 31.6%). In logistic regressions, the number of distinct medications was the best predictor of one-year and two-year risks of emergency department use (c = 0.653, 0.654), but the index-year healthcare expenditures performed the best in predicting one-year and two-year risks of hospitalizations (c = 0.684, 0.676) and high-expenditure cases (c = 0.810, 0.823). The updated CCI consistently outperformed the original CCI in predicting the outcomes of interest. Conclusions: In a diabetic population under age 65, the number of distinct medications and baseline healthcare expenditures appeared to have superior or similar powers compared to the CCI or CDS for the prediction of future healthcare utilization and expenditures. The updated CCI was a better predictor than the original CCI in this population.