Machine learning-based prediction of bronchiolitis in children under two years: a multicenter study within a single metropolitan area.

BACKGROUND

Bronchiolitis is a leading cause of hospitalization in infants, yet early and accurate risk prediction remains a clinical challenge. Traditional models often lack the complexity to capture nonlinear interactions or the transparency required for clinical trust.

OBJECTIVE

This multicenter study aimed to develop, validate, and interpret machine learning (ML) models for predicting bronchiolitis risk in children under two years of age using clinical and socioeconomic determinants.

METHODS

In this retrospective study, data from 1,260 children (529 with bronchiolitis and 731 without) were collected at 5 pediatric centers within a single metropolitan area, from January 2023 through December 2024. Seventeen predictors were analyzed. After rigorous preprocessing and K-NN imputation, eight ML algorithms, including four base learners and four ensemble methods, were developed. Model performance was evaluated using AUC and metrics derived from confusion matrices. Clinical utility was assessed via Decision Curve Analysis (DCA) and calibration plots. Explainable AI (XAI) was implemented using Shapley Additive exPlanations (SHAP) to interpret model decisions.

RESULTS

XGBoost achieved an AUC of 0.863 [0.828-0.897] in the test set, demonstrating higher predictive performance than base learners such as NB (AUC: 0.583 [0.529-0.638]). DCA and calibration plots confirmed that XGBoost provides superior net benefit and calibration performance across a range of threshold probabilities. SHAP analysis identified overcrowding, severe acute malnutrition, and maternal smoking as the most influential predictors, highlighting the critical role of socioeconomic and environmental factors in disease susceptibility.

CONCLUSION

The integration of XGBoost with SHAP values provides a robust, transparent, and clinically actionable framework for early prediction of bronchiolitis based on variables available at initial clinical presentation. By identifying high-risk infants through interpretable AI, this model supports early triage and optimized resource allocation, fostering the transition toward precision pediatric medicine.

CLINICAL TRIAL NUMBER: Not applicable.