Advancements in machine learning may offer clinicians a safer, faster way to distinguish patients with true peanut allergies from those who are sensitized yet able to tolerate peanuts. According to a newly published study in the Journal of Allergy and Clinical Immunology: Global, researchers have developed artificial intelligence (AI) models that predict oral food challenge outcomes with high accuracy, potentially helping clinicians determine when an in-clinic food challenge is necessary.
Oral food challenges (OFCs) remain the gold standard for diagnosing peanut allergy. During an OFC, patients ingest escalating doses of peanut under medical supervision to determine whether they experience an allergic reaction. Although considered the most definitive diagnostic method, OFCs are resource-intensive, time-consuming, and carry the risk of triggering severe allergic reactions.
To address this clinical dilemma, investigators from Charité–Universitätsmedizin Berlin evaluated whether machine learning could integrate routinely collected clinical and immunological information to predict challenge outcomes. The study examined data from 96 peanut-sensitized individuals ranging from 1 to 56 years of age, including 74 patients with confirmed peanut allergies and 22 who were sensitized but able to tolerate peanut.
The research team developed and compared three types of machine learning models: one combining clinical data with results from the basophil activation test (BAT), one relying solely on clinical data, and a third using only BAT measurements. The combined model performed best, correctly predicting OFC outcomes with 96% accuracy during internal cross-validation. Notably, however, the clinical-only model performed nearly as well at 95% accuracy, while the BAT-only model achieved 83%.
That finding could prove particularly important because the clinical-only approach came within a single percentage point of the model incorporating BAT results. If confirmed in larger studies, a highly accurate model based on clinical information alone could potentially offer a more practical approach without requiring the additional specialized laboratory testing involved in BAT.
Using explainable AI techniques, the researchers were also able to determine which measurements contributed most to the models’ predictions. Levels of immunoglobulin E (IgE) antibodies directed against Ara h 2 — a major peanut allergen component — were particularly influential. Higher Ara h 2-specific IgE concentrations and larger reactions during peanut skin testing were associated with allergic responses during OFCs, while measurements from the BAT, which assesses how immune cells react when exposed to peanut allergens, also contributed to the predictions.
The researchers then tested the models using an independent cohort from the LEAP (Learning Early About Peanut Allergy) study. In this external validation, the combined clinical and BAT model correctly classified 86% of participants, while the clinical-only model achieved 85% accuracy, supporting the potential for the approach to work beyond the original study population.
The technology was less successful when researchers attempted to predict more detailed aspects of an allergic reaction. Separate regression models designed to estimate the maximum amount of peanut a patient could tolerate and the severity of a potential reaction performed substantially worse than the models designed simply to predict whether an OFC would result in an allergic reaction.
The researchers describe their findings as a proof of concept rather than a replacement for oral food challenges. If validated, explainable machine learning could eventually help allergists distinguish peanut-allergic patients from those who are merely sensitized and help determine when an OFC is necessary. Prospective studies involving larger, more diverse populations across multiple medical centers will be required before such models can be incorporated into routine allergy care.
