Machine learning-based prediction of food sensitization in infants and toddlers with atopic dermatitis.

Abstract

Background: Food sensitization (FS) is a common comorbidity in children with atopic dermatitis (AD), often leading to the atopic march. Identifying high-risk patients in primary care is crucial to prioritize specialist referrals and prevent unnecessary elimination diets.

Objective: To develop and validate a machine learning (ML) based clinical decision support tool to predict IgE-mediated FS risk in infants and toddlers (0-3 years) diagnosed with AD using routine clinical and laboratory parameters.

Methods: We conducted a retrospective cohort study of children diagnosed with AD. Patients were classified as FS positive or FS negative based on specific IgE levels and skin prick tests. Ten ML algorithms were trained and compared using 5-fold stratified cross-validation. Feature importance was analyzed, and an interactive web-based tool was developed for clinical application.

Results: Among 340 patients, 156 (45.8%) had IgE-mediated FS. Key independent predictors of FS included higher total IgE levels, younger age at presentation, elevated eosinophil counts, and breastfeeding at the time of symptom onset. Boosting-based models achieved superior performance; specifically, the AdaBoost model demonstrated high stability and discrimination (ROC AUC: 0.815 in CV and 0.819 in the held-out test set). By optimizing the probability threshold to 0.40, the model achieved a recall of 92% and an F1-score of 0.75, effectively minimizing false negatives.

Conclusion: The ML model may serve as a screening and risk-stratification tool for IgE-mediated FS in children with AD. By utilizing routine laboratory data, this web-integrated tool can support primary care physicians in risk stratification, ensuring timely referral for high-risk cases while reducing unnecessary testing in low-risk infants.