Evaluation of the Discriminative Ability of Multiple Comorbidity Indices

Taiwan’s National Health Insurance Research Database (NHIRD) has been increasingly used to assess health services utilization and medical care outcomes of patients. Researchers have recognized the importance of adequately adjusting for comorbidities in studies. However, many researchers in Taiwan adopt the Charlson comorbidity index (CCI), rather than other comorbidity indices that have demonstrated better performance for predicting health care utilization or medical care outcomes in the literature, such as the Roman index or the Elixhauser index. Is the phenomenon due to some unique characteristics of people or health care system in Taiwan? The study aims to explore the issue by comparing the discriminatory power of various comorbidity indices by using the NHIRD. More specifically, the current research will compare the performance of predicting health services utilization and medical care outcomes between two categories of comorbidity measures: diagnosis-based comorbidity indices versus medication-based comorbidity indices. Data will be sourced from the Longitudinal Health Insurance Database (LHID), two-million beneficiaries, the third cohort. The study population comprised all patients who were hospitalized due to chronic obstructive pulmonary disease (COPD) for the first time. Five diagnosis-based comorbidity indices (the CCI, the Deyo index, the Romano index, the D`Hoore index, and the Elixhauser index) and two medication-based comorbidity indices (chronic disease score and the RxRisk-V model) were compared regarding the performance of predicting medical expenditures of subjects within one year after discharge and in-hospital mortality. The baseline model included patient’s age and gender, if surgery undertaken when hospitalized, and the length of stay. Comorbidity measures were analyzed both as an aggregate index (the weighted sum of comorbidity indicators) and as individual components (simply assigning an indicator depending on whether the comorbidity was present). Three target years of data were compared. The discriminatory power of the comorbidity measures was assessed using the c-statistics derived from multiple logistic regression models.