Improving hospital readmission prediction using individualized utility analysis.

Area of Science:

• Health Informatics
• Machine Learning Applications
• Healthcare Economics

Background:

• Machine learning (ML) models for readmission mitigation are often chosen based on predictive accuracy.
• This approach may not align with maximizing the utility of limited healthcare resources.
• A gap exists in understanding how different ML model selection criteria impact resource allocation effectiveness.

Purpose of the Study:

• To compare the utility and discriminative ability of ML models for allocating readmission-mitigation interventions.
• To determine if models optimized for utility offer greater value than those optimized for discrimination.
• To inform best practices for selecting ML models in healthcare resource allocation.

Main Methods:

• Retrospective utility analysis of ML models using claims data from 513,495 commercially-insured inpatients.
• Evaluation of models trained to predict readmissions versus those trained to predict cost.
• Calculation of estimated utility gain (reduction in 90-day cost) and Area Under the Curve (AUC) for each model.

Main Results:

• A Gradient Boosting Decision Tree (GBDT) model for readmission prediction yielded $104 utility gain per patient (AUC 0.76).
• A model predicting cost as a proxy achieved higher utility ($175.94 per patient) but lower discrimination (AUC 0.62).
• Hybrid models showed comparable performance; optimal model choice depends on intervention cost and efficacy.

Conclusions:

• ML models can be ranked differently based on utility versus discriminative ability.
• Directly optimizing ML models for utility is crucial for effective healthcare resource allocation.
• Future ML model selection should consider overall utility to maximize benefits.