Ensemble Machine Learning Models for Predicting Patients With High Usage: Model Validation and Economic Impact...

Area of Science:

• Health Informatics
• Machine Learning
• Predictive Analytics

Background:

• Machine learning models are increasingly utilized for predicting high healthcare utilization.
• Early identification of at-risk patients enables targeted interventions.

Purpose of the Study:

• Evaluate the predictive performance of multiclass ensemble models for healthcare usage.
• Assess the economic impact of these models in real-world scenarios.

Main Methods:

• Four binary classification models (boosted trees, MARS, MLP, logistic regression) were extended using a stacking ensemble approach.
• Multiclass models predicted length of stay (LOS) and emergency department (ED) visits across defined strata.
• Ensemble models were trained on 2020-2021 data and validated on 2021-2022 data using AUC, accuracy, and confusion matrix metrics.

Main Results:

• Boosted tree ensemble models demonstrated the highest performance, achieving AUC scores of 0.6877 for LOS and 0.7601 for ED visits.
• Models correctly classified 30.3% of inpatient LOS and 39.8% of ED visits, identifying 77% of future inpatient and 73.9% of future ED users.
• Economic analysis projected an average cost reduction of US $111 million with the boosted tree model using logistic regression base learners.

Conclusions:

• Multiclass ensemble models effectively predict multilevel healthcare usage.
• These models offer potential for significant cost savings and support targeted interventions.
• The findings can inform planning and budgeting for population health programs, particularly for conditions like diabetes.