A hospital wide predictive model for unplanned readmission using hierarchical ICD data

Area of Science:

• Health Informatics
• Clinical Data Science
• Predictive Modeling in Healthcare

Background:

• Hospitals generate vast administrative, billing, and registration data.
• Secondary use of patient data can improve care with minimal extra cost.
• Unplanned readmissions are a key metric for quality improvement programs.

Purpose of the Study:

• To explore the potential of secondary patient data for predicting unplanned readmissions.
• To model administrative, billing, and International Classification of Diseases (ICD) hierarchical data.
• To develop a decision support tool for hospital quality improvement.

Main Methods:

• Single-center, hospital-wide observational cohort study of 29,702 adult patients discharged in 2016.
• Compared logistic regression, penalized logistic regression, gradient boosting, and Random Forests.
• Evaluated model performance using Area Under the ROC Curve (AUC), investigating ICD hierarchy levels.

Main Results:

• Random Forests models achieved the best predictive performance.
• An AUC of 0.77 was obtained using Random Forests with diagnosis category and procedure code.
• This model showed a 7% improvement over baseline logistic regression.

Conclusions:

• Random Forests models incorporating the ICD hierarchy, particularly diagnosis categories, accurately predict unplanned readmissions.
• This approach enhances interpretability and reduces the number of predictor variables.
• The model's performance supports its use as a clinical decision support tool.