Improving hospital quality risk-adjustment models using interactions identified by hierarchical group lasso...

Area of Science:

• Health Services Research
• Biostatistics
• Health Informatics

Background:

• Risk-adjustment (RA) models are crucial for comparing patient outcomes across hospitals by accounting for illness severity.
• Traditional RA models often overlook interaction effects or use stratification, which can be problematic with rare events and sparse data.
• Existing Agency for Healthcare Research and Quality (AHRQ) hospital quality indicators have limitations in performance and interpretability.

Purpose of the Study:

• To develop and evaluate a novel method for identifying clinically meaningful interactions in risk-adjustment models.
• To improve the performance and interpretability of hospital quality indicators by incorporating interaction effects.
• To address limitations of current RA models, particularly those using stratification with sparse data.

Main Methods:

• Utilized de-identified patient discharge data from 14 State Inpatient Databases.
• Applied hierarchical group lasso regularization (HGLR) to identify first-order interactions in AHRQ inpatient quality indicators (IQI 09, IQI 11) and Patient Safety Indicator 14 (PSI 14).
• Compared HGLR models against stratum-specific and composite main effects models using covariates selected by least absolute shrinkage and selection operator (LASSO).

Main Results:

• HGLR successfully identified clinically significant synergistic and antagonistic interactions for all tested AHRQ quality indicators.
• Identified interactions, such as between hypertension and respiratory failure in IQI 11, highlight patient subgroups needing special perioperative attention.
• HGLR-selected features resulted in composite models with similar or superior performance compared to LASSO-selected features.

Conclusions:

• HGLR offers a scalable and customizable approach to identify important interactions, maintaining or improving RA model performance in heterogeneous risk populations.
• This method overcomes limitations of stratified models on sparse data, leading to improved model calibration and reduced bias.
• HGLR is valuable for hospitals and policymakers using RA models for public reporting and payment programs, enhancing the accuracy of quality assessments.