The measures of clinical status used to predict costs must pay the most attention possible to medical conditions and clinical complexity. Length of stay (LOS), which has been used as a proxy for resource consumption, is not a direct measure of costs. Classification and regression trees, which are used in defining iso-resource groups, can be affected by overfitting and are based on a priori choices of the splitting attributes. Finally, current approaches are mainly concerned in estimating average group costs and do not attempt to estimate individual case costs.
We sought to define comprehensive measures of clinical status and detailed measures of resource consumption. We also sought to predict individual inpatient rehabilitation costs through multiple regression models.
A prospective analysis was conducted of all rehabilitation cases admitted to 5 Italian inpatient facilities during a period of 12 months. All admissions underwent repeated Minimum Data Set–Post Acute Care (MDS-PAC) schedules to collect information on clinical status and treatment provided. We used factorial analysis to yield continuous variables representing clinical characteristics, and we priced treatments to obtain cost of stay. We used linear regression models to predict cost of stay and validated the model-based cost predictions by data-splitting.
We collected 9720 MDS-PAC schedules from 2702 hospital admissions. The multivariate regression models fitted costs reasonably well with r2 values of at least 0.34. On cross-validation, the ability of the regression models to predict cost was confirmed.
We were able to estimate actual rehabilitation costs and define reliable regression models to predict costs from individual patient characteristics. Our approach identifies the contribution of any single patient characteristic to rehabilitation cost and tests the assumptions of the analysis.