AI Model Shows Promise for Predicting Peanut Allergy

MACHINE LEARNING could help clinicians distinguish patients with peanut allergy from those who are sensitised but able to tolerate peanuts, potentially reducing reliance on oral food challenges (OFCs).

A new study has developed explainable artificial intelligence (AI) models using clinical and immunological data from peanut-sensitised children and adults. The best-performing model achieved 96% accuracy in cross-validation and maintained 86% accuracy when evaluated using an independent cohort.

Machine Learning Targets a Diagnostic Challenge

OFCs remain the gold standard for diagnosing peanut allergy, requiring patients to consume increasing doses of peanut under medical supervision. However, they are resource-intensive and carry a risk of severe allergic reactions.

Researchers from Charité–Universitätsmedizin Berlin, Germany, investigated whether machine learning could combine routinely collected clinical and immunological measures to predict OFC outcomes.

The study included 96 peanut-sensitised participants aged 1- 56 years. 74 of them had a confirmed peanut allergy and 22 who were sensitised but tolerant. Researchers developed models using clinical data, basophil activation test (BAT) results, or a combination of both.

Combined Testing Improved Prediction of Peanut Allergy

The researchers developed three types of machine learning model to predict whether peanut-sensitised participants would have an allergic reaction during an oral food challenge. Models either combined clinical and basophil activation test (BAT) data, used clinical data alone, or relied solely on BAT measurements.

The model combining clinical and BAT data performed best, correctly predicting OFC outcomes with 96% accuracy during cross-validation. However, a model based on clinical data alone performed similarly, with 95% accuracy, while the BAT-only model achieved 83%.

Explainable AI analysis also allowed the researchers to identify which measurements contributed most to these predictions. Levels of IgE antibodies against Ara h 2, a major peanut allergen, were particularly influential. Higher Ara h 2-specific IgE levels and larger reactions during peanut skin testing were associated with an allergic response during an OFC. Measures from the BAT, which assesses how immune cells respond when exposed to peanut allergens, also contributed to the predictions.

Importantly, the researchers also evaluated the models using data from another independent LEAP study. The combined model correctly classified 86% of participants in this separate dataset, while the clinical-only model achieved 85% accuracy, supporting the potential for the approach to work beyond the original study population.

Further Validation Is Needed

The approach was less successful at predicting maximum tolerated peanut dose and reaction severity, with both regression models designed to make more detailed predictions about reactions performing considerably worse than the binary allergy prediction models.

The researchers describe the findings as proof of concept, suggesting that explainable machine learning could eventually support clinicians in distinguishing peanut-allergic patients from sensitised but tolerant individuals and help determine when OFCs are necessary.

However, prospective validation in larger, diverse, multicentre populations will be needed before the models can be incorporated into routine allergy care.

Reference

Kemmler P E. et al. Leveraging Explainable AI on Clinical Data from Oral Food Challenges to Predict Peanut Allergy. J Allergy Clin Immunol Glob. 2026;100770.

Featured image: Daisy Daisy on AdobeStock