Machine Learning Readmission Risk Modeling: A Pediatric Case Study.

Area of Science:

• Pediatric healthcare analytics
• Machine learning in medicine
• Healthcare cost optimization

Background:

• Hospital readmission prediction in pediatric settings is under-researched.
• Existing studies focus on frequency analysis, lacking predictive modeling.
• Predictive models can identify preventable readmissions and reduce healthcare costs.

Purpose of the Study:

• To evaluate machine learning techniques for predicting all-cause readmissions.
• Focus on the emergency department of a pediatric hospital in Santiago, Chile.
• Assess predictive performance for 30-day readmissions.

Main Methods:

• Retrospective analysis of a six-year pediatric admissions dataset.
• Formulated readmission prediction as a binary classification problem.
• Employed data preprocessing, class imbalance correction (SMOTE), and repeated cross-validation (RCV).

Main Results:

• SMOTE significantly improved recall for class imbalance correction.
• Naive Bayes (NB) achieved the highest Area Under the Curve (AUC) of 0.65.
• Shallow multilayer perceptron showed the best Precision-Positive Predictive Value (PPV) and F-score; NB and Support Vector Machines (SVM) performed comparably.

Conclusions:

• Recommends Naive Bayes (NB) with a Gaussian distribution model for pediatric readmission prediction.
• NB demonstrated robustness across various training dataset sizes.
• The developed approach can aid in identifying and preventing costly readmissions.