Related Experiment Videos
Derivation and Validation of Charlson, Elixhauser and RxRisk Cost-Based Comorbidity Indices: A Register-Based Study
Camilla Nystrand1,2, Naimi Johansson3, Johannes Blom3,4
1Department of Learning, Informatics, Management and Ethics, Karolinska Institutet, Tomtebodavägen 18 A, Solna, 171 77, Stockholm, Sweden. camilla.nystrand@ki.se.
Background:
Comorbidity indices are widely applied to adjust for confounding in observational research. However, health economic analyses often adjust for comorbidity using mortality-based indices, despite health conditions being likely associated differently, both in terms of magnitude and direction, to resource use than to mortality.
Objective:
We aimed to derive outcome-specific cost-based indices based on the Charlson, Elixhauser and RxRisk comorbidity measures, and to internally validate their performance for predicting annual healthcare costs compared with published mortality-based weights.
Methods:
Using a retrospective register-based design with 11 years of data, diagnostic (International Classification of Diseases, Tenth Revision) and Anatomical Therapeutic Chemical (prescribed drugs) codes determined whether 350,000+ individuals had one or several of the disease categories defined by each comorbidity index. Outcomes were annual total healthcare costs (primary, specialised and inpatient care, prescribed drugs) and inpatient costs. Prediction models were two-part generalised linear models, adjusting for age in splines, sex and year fixed effects. Internal validation with repeated cross-validation assessed predictive performance. Calibration curves with bootstrapped variance were plotted, and improvement in prediction between mortality- and cost-based indices was assessed.
Results:
The cost-based RxRisk index was better in terms of calibration and performed significantly better than the original mortality-based RxRisk in predicting total healthcare and inpatient costs. The mortality- and cost-based Charlson models showed inconclusive results in relation to prediction accuracy, while well calibrated. The cost-based Elixhauser index reduced prediction errors and significantly improved explained variance compared with the mortality-based index. Decile calibration plots illustrate the superior comparative advantage of RxRisk across groups, where observed versus predicted mean cost deviation averaged 7% across deciles using the cost-based RxRisk, whereas the Charlson and Elixhauser indices deviated 32-40%.
Conclusions:
In future research, when using a comorbidity index to adjust for confounding in cost analyses, our results suggest using the cost-based RxRisk index.