Incorporation of expert knowledge in the statistical detection of diagnosis related group misclassification

Abstract

Background: In activity based funding systems, the misclassification of inpatient episode Diagnostic Related Groups (DRGs) can have significant impacts on the revenue of health care providers. Weakly informative Bayesian models can be used to estimate an episode's probability of DRG misclassification.

Methods: This study proposes a new, Hybrid prior approach which utilises guesses that are elicited from a clinical coding auditor, switching to non-informative priors where this information is inadequate. This model's ability to detect DRG revision is compared to benchmark weakly informative Bayesian models and maximum likelihood estimates.

Results: Based on repeated 5-fold cross-validation, classification performance was greatest for the Hybrid prior model, which achieved best classification accuracy in 14 out of 20 trials, significantly outperforming benchmark models.

Conclusions: The incorporation of elicited expert guesses via a Hybrid prior produced a significant improvement in DRG error detection; hence, it has the ability to enhance the efficiency of clinical coding audits when put into practice at a health care provider.

Keywords: Bayesian analysis; Clinical coding; DRGs; Health informatics; Statistical modeling.