Predicting Pediatric Asthma Readmissions Through Machine Learning: Performance With a National Administrative Database

BACKGROUND: Asthma remains a leading cause of preventable pediatric hospitalizations and is largely socially-mediated. Identification of children with asthma at risk of hospital readmission through machine learning (ML) could streamline social care interventions tailored to individual needs, fostering precision social medicine (PSM). OBJECTIVE: To develop and externally validate a ML model for predicting 180-day asthma readmissions and to compare its performance with a regression model. METHODS: Our study included encounters for children aged 4-18 years hospitalized with asthma exacerbations in the Pediatric Health Information Systems database from 2016 to 2024. Models were trained on 36 hospitals and validated on an unseen set of 11 hospitals. The primary outcome was readmission within 180 days. We compared a regression model (GLM) and an XGBoost (XGB) model. Performance was assessed using the area under the receiver operating characteristic curve (ROC AUC) and area under the precision-recall curve (PR AUC). RESULTS: There were 137,854 encounters in the training set (outcome prevalence: 10.1%) and 35,876 encounters in the test set (outcome prevalence: 9.8%). The XGB model achieved superior performance compared with the GLM model (ROC AUC: 0.716 [95% CI: 0.707, 0.726] vs 0.702 [95% CI: 0.692, 0.712]; PR AUC: 0.270 [95% CI: 0.256, 0.285] vs 0.250 [95% CI: 0.236, 0.264]; p < 0.001 for both). CONCLUSIONS: ML achieved moderate performance with modestly improved prediction compared with conventional regression. Integration of outpatient, pharmacy, and individual-level social data that comprehensively reflects children's health may further enhance the ability of ML to enable PSM approaches.