Regression tree construction by bootstrap: model search for DRG-systems applied to Austrian health-data

Area of Science:

• Health economics
• Medical informatics
• Statistical modeling

Background:

• Diagnosis Related Group (DRG) systems allocate hospital resources based on performance.
• Current statistical models (regression trees) often require manual adjustments for medical and ethical reasons.
• Manual adaptations can negatively impact model performance.

Purpose of the Study:

• To systematically search for alternative regression trees beyond manual adjustments.
• To develop a robust method for constructing diverse and accurate models for DRG systems.
• To propose a two-step model selection process balancing complexity and clinical relevance.

Main Methods:

• Utilized the bootstrap-based bumping method to generate diverse regression tree models.
• Implemented a two-step model selection: first, determining optimal complexity, then selecting a medically sound and accurate model.
• Analyzed 8 Austrian DRG datasets to evaluate model diversity and accuracy.

Main Results:

• Bootstrap-based trees demonstrated superior predictive accuracy compared to the CART algorithm.
• Diverse models were successfully constructed even with small tree sizes.
• These diverse models were as accurate or more accurate than single CART-generated models.

Conclusions:

• Bumping is an effective technique for creating diverse, accurate regression trees for DRG systems.
• The proposed method enhances candidate model selection for DRG resource allocation.
• Bumping and the selection approach are applicable to other medical decision-making and prognosis tasks.