Prediction of future onset of allergic rhinitis and analysis of risk factors in children with food allergy during infancy.

The atopic march describes the progression from infantile food allergy (FA) to allergic rhinitis (AR) in later childhood. However, not all children with FA follow this trajectory, and predictors for AR development in this particular cohort remain poorly characterized. This study aims to identify factors influencing AR development in infants with FA and construct a predictive model for clinical application. A prospective cohort included 447 infants (0-2 years) with positive food allergens, followed up to 6 years with questionnaire/clinical data. After preprocessing, feature selection was performed using variance inflation factor (VIF) and Lasso regression. Five machine learning models (logistic regression, SVC, random forest, XGBoost, and KNN) were trained and validated using stratified fivefold cross-validation and an independent test set. In cross-validation, logistic regression achieved the highest mean AUC of 0.9403 (95% CI 0.9040-0.9766), and SVC obtained the best F1 score (0.9341). On the test set (n = 90), logistic regression maintained the highest AUC (0.8466), while random forest achieved a recall of 1.0. The top predictive features identified were parental allergy history, frequency of antibiotic use, cephalosporin use, and daily duration of tobacco smoke exposure.

conclusionThe machine learning prediction model, particularly logistic regression, shows good practical value for early identification of FA infants at high risk of developing AR. Early avoidance of the identified modifiable risk factors may aid primary and secondary prevention. WHAT IS KNOWN: • The atopic march from food allergy (FA) to allergic rhinitis (AR) is well-recognized, yet outcomes are heterogeneous and pediatricians lack validated risk tools for infants with FA. WHAT IS NEW: • This study identifies three easily assessable predictors (parental AR history, tobacco exposure duration, and antibiotic use frequency) and validates a Multinomial Naive Bayes model (AUC=0.915) that accurately stratifies AR risk in FA children, significantly outperforming traditional regression.