Building interpretable predictive models for pediatric hospital readmission using Tree-Lasso logistic regression.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Modeling

Background:

• Unplanned pediatric readmissions pose challenges to healthcare quality.
• Electronic health data present complexities like high dimensionality, sparsity, and class imbalance for predictive modeling.
• Interpretable models are crucial for real-world clinical application and actionable insights.

Purpose of the Study:

• To develop accurate and interpretable predictive models for pediatric readmission risk.
• To integrate data-driven sparse logistic regression with domain knowledge from the International Classification of Diseases 9th Revision Clinical Modification (ICD-9-CM) hierarchy.
• To propose a method for quantifying model interpretability and assessing solution stability.

Main Methods:

• Analysis of over 66,000 pediatric hospital discharge records (2009-2011).
• Utilized a Tree-Lasso regularized logistic regression model incorporating ICD-9-CM hierarchy for interpretability.
• Compared the Tree-Lasso approach with traditional Lasso logistic regression.

Main Results:

• The Tree-Lasso model demonstrated comparable prediction accuracy (AUC ~0.78) to traditional Lasso.
• Integration with ICD-9-CM hierarchy resulted in more interpretable models focused on high-level diagnoses.
• Model interpretations align with existing medical understanding of pediatric readmission, with lower information loss than Lasso.

Conclusions:

• A novel method combining domain knowledge (ICD-9-CM) and sparse algorithms (Tree-Lasso) enhances predictive model interpretability for readmission risk.
• The developed models are applicable to general pediatric populations and subpopulations, with interpretations consistent with medical knowledge.
• Quantitative assessment of interpretability goes beyond simple feature counts, offering deeper insights.