Ensemble model-based prediction of school-age asthma risk in children with food allergy

Early-life food allergy (FA) is a major childhood asthma risk, but prospective screening tools for high-risk infants aged 0-2 are lacking, delaying early intervention.

To develop an ensemble-learning model with cost-sensitive optimization for early asthma risk screening in FA infants (0-2 years) and provide quantitative clinical decision support.

This prospective cohort enrolled 447 FA-confirmed infants (0-2 years), collecting 135 baseline predictors across 7 categories. After median 4.5-year follow-up to school age (≥6 years), we generated 3 clinically interpretable features and built a two-stage weighted ensemble (Random Forest, XGBoost, LightGBM, Deep Residual Networks) with cost-sensitive learning (missed-to-false-alarm ratio = 5:1) to optimize thresholds.

On the independent test set, AUC = 0.886, recall = 1.00 (zero missed). Decision-curve analysis showed superior net benefit across broad thresholds, including optimal cutoff 0.09. Top predictors: RSV infection, wheeze frequency, nocturnal sleep, antibiotic exposure, and rhinitis duration. Sensitivity analysis confirmed stable zero missed at cost ratios ≥5:1, indicating 5:1 was not the key driver.

The model preliminarily predicts school-age asthma risk in FA infants. Its high-recall design avoids missed diagnoses and offers quantitative references for early intervention and individualized follow-up before age 2.

Core finding: A performance-weighted ensemble model predicts school-age asthma in food-allergic infants (0-2 years) with an AUC of 0.886 and recall 1.00 using only clinical data. Contribution: A cost-sensitive, feature-engineered model tailored to this high-risk population; identifies RSV, wheeze grading, and antibiotic load as top predictors.

Enables early, non-invasive screening before age 2, avoids missed high-risk children, and supports targeted preventive interventions in primary care.