Leveraging machine learning models to forecast pediatric allergic rhinitis exacerbation risk based on environmental...

Acute exacerbations of pediatric allergic rhinitis (AR) are difficult to anticipate, and individualized risk tools integrating clinical and environmental data are lacking. We developed an interpretable machine-learning model to predict 90-day AR exacerbation risk in children and conducted a preliminary transportability assessment in an independent cohort, following TRIPOD + AI guidelines.

A development cohort (n = 2,000) was assembled by linking NHANES (2005-2006 and 2011-2018 cycles) with EPA Air Quality System and NOAA meteorological data; an independent external cohort (n = 50) was recruited at Wuhan Children's Hospital (2023-2024). Clinical predictors [Total Nasal Symptom Score (TNSS), total IgE, eosinophils, comorbidities] and environmental exposures (same-day and 0-7-day-lag PM_2.5, PM₁₀, O₃, NO₂, temperature, humidity) were analyzed. L2-regularized logistic regression (LR) was prespecified as the primary model and benchmarked against random forest, XGBoost, LightGBM, and SVM using nested 5-fold cross-validation and multiple imputation (m = 5). Discrimination, calibration, decision-curve analysis, and SHAP interpretability were evaluated.

Exacerbation occurred in 34.4% (687/2,000) of the development cohort and 36.0% (18/50) of the external cohort. The primary LR model achieved an internal AUC of 0.81 (95% CI: 0.74-0.88) and an external AUC of 0.71 (95% CI: 0.50-0.92); the wide confidence interval reflects the limited precision of the small validation sample. Baseline TNSS, same-day PM_2.5, and total IgE were the most influential predictors across feature-selection and SHAP analyses. Children exposed to PM_2.5 > 75 μg/m³ had approximately five-fold higher exacerbation odds than those exposed to <35 μg/m³ (OR = 5.08, 95% CI: 3.45-7.47). A sensitivity model excluding TNSS retained moderate discrimination (external AUC 0.66, 95% CI: 0.49-0.83), supporting the independent contribution of environmental and immunologic factors. Decision-curve analysis showed positive net benefit across threshold probabilities of 0.10-0.35.

An interpretable LR model integrating baseline symptom burden with ambient PM_2.5 exposure demonstrated promising predictive performance for pediatric AR exacerbation. However, the small external cohort (n = 50) yielded wide confidence intervals, and findings should be regarded as preliminary transportability evidence rather than definitive external validity. Larger multicenter prospective studies with site-specific recalibration are required before clinical implementation.