Machine learning for hospital readmission prediction in pediatric population

Area of Science:

• Utilizes advanced machine learning (ML) techniques for predictive modeling in healthcare.
• Focuses on pediatric readmission risk assessment within a tertiary care setting.

Background:

• Pediatric readmissions pose significant burdens on patients, families, and healthcare systems.
• Accurate identification of high-risk patients is crucial for effective resource allocation and intervention.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting potentially avoidable 30-day readmissions in pediatric patients.
• To identify key clinical and demographic factors associated with increased readmission risk.

Main Methods:

• Retrospective cohort study of 9,080 pediatric patients admitted to a tertiary university hospital.
• Six ML algorithms (CART, RF, GBM, XGBoost, Decision Tree, LR) were applied to a training/testing dataset (75%/25%).
• Model performance was assessed using AUC, sensitivity, specificity, and Youden's J-index.

Main Results:

• The rate of avoidable 30-day readmissions was 9.5%.
• XGBoost, Random Forest, GBM, and CART showed comparable performance (AUC).
• XGBoost with bagging imputation achieved the highest Youden's J-index (0.484) with an AUC of 0.814.
• Key predictors included cancer diagnosis, age, red blood cell count, leukocytes, red cell distribution width, sodium levels, elective admission, and multimorbidity.

Conclusions:

• Machine learning, specifically XGBoost, demonstrates strong potential for predicting 30-day pediatric readmissions.
• Implementation in hospital systems can facilitate early risk identification and targeted interventions.
• The model aids in optimizing healthcare strategies for pediatric readmission prevention.