AI Improves Food Allergy Diagnosis Accuracy | Healio

Philadelphia – Advances in machine learning are offering a potential path toward more accurate and less invasive food allergy diagnoses, according to research presented at the American Academy of Allergy, Asthma & Immunology (AAAAI) Annual Meeting. The study, focused on peanut allergy in young children, suggests that integrating multiple biomarker datasets with deep learning models could significantly improve diagnostic precision compared to current standard practices, potentially reducing the need for oral food challenges – a procedure carrying inherent risks.

Augmenting, Not Replacing, Clinical Judgment

Researchers led by McKenzie Williams, a Howard University Karsh STEM Scholar, alongside Kyung Won Lee, PhD, and Mayte Suarez-Farinas, PhD, of Icahn School of Medicine at Mount Sinai, and Carla M. Davis, MD, of Howard University College of Medicine, emphasized that these models are intended to augment, not replace, the expertise of clinicians. The goal is to provide probabilistic risk estimates based on comprehensive data analysis, leading to safer and more informed diagnostic decisions. “Providing probabilistic risk estimates grounded in large-scale data could make diagnosis safer and more precise,” the researchers stated.

Currently, diagnosing food allergies often relies on a combination of skin prick tests, allergen-specific IgE blood tests, and, oral food challenges (OFCs) – considered the gold standard but carrying the risk of severe reactions. However, these traditional methods can struggle to differentiate between true allergy and mere sensitization, leading to overdiagnosis, unnecessary dietary restrictions, and anxiety for families. The study aimed to address these limitations by exploring whether machine learning could better integrate existing biomarker data to predict allergic reactions.

How the Study Worked

The research team analyzed data from 146 OFCs conducted on children aged one to four years as part of the IMPACT trial. The dataset included results from skin prick tests, allergen-specific IgE levels, and measurements of serum component proteins related to peanut allergy (peanut-IgE and peanut-IgG4 rAra h 1, 2, 3 and 6), as well as data from bead-based epitope assays (BBEA). They then trained both machine learning and deep learning convolutional neural networks to assess how well these models could predict OFC outcomes compared to current diagnostic criteria and against each other. Oral food challenges remain a critical, though potentially risky, component of allergy diagnosis.

The current diagnostic criteria used in the study involved a skin prick test result of 4mm or greater, a peanut-specific IgE level of 0.35 or higher, and standard thresholds for serum component proteins. Researchers found that these traditional thresholds demonstrated only modest accuracy, with area under the curve (AUC) values ranging from 48% to 61%.

Machine Learning Shows Promise

The study revealed that machine learning models significantly improved diagnostic accuracy, achieving approximately a 40% improvement over current diagnostic criteria. For example, a machine learning model trained on serum IgE data achieved an AUC of 76%, while a model incorporating both serum IgE and serum component proteins reached an AUC of 85%. This suggests that combining multiple biomarkers provides a more comprehensive picture of a patient’s allergic risk.

However, the most substantial gains were observed with deep learning models. These models, which are capable of identifying more complex patterns in data, demonstrated a 10-15% improvement in AUC, a 35% improvement in sensitivity, and a 12-19% improvement in positive predictive value compared to machine learning models. Specifically, a deep learning model integrating serum IgE with component-resolved peanut markers achieved an impressive AUC of 94%, with high sensitivity (88.9%) and specificity (84.5%).

The Power of Integrated Data

The researchers emphasized that no single biomarker was sufficient on its own to accurately predict allergic reactions. The predictive power stemmed from the patterns observed across multiple biomarkers. “The most striking finding was that no single biomarker was sufficient on its own,” the researchers explained. “The predictive power came from the pattern across markers. This reinforces the idea that food allergy is biologically complex and that AI is particularly well suited to modeling that complexity.”

This finding underscores the importance of a holistic approach to food allergy diagnosis, moving beyond evaluating biomarkers in isolation. The study highlights the potential of AI to uncover subtle relationships between different biomarkers that might be missed by traditional statistical methods.

What’s Next for AI-Driven Allergy Diagnosis?

The researchers outlined three key areas for future research: validation, transparency, and implementation. First, they stressed the need for larger, more diverse datasets to ensure that the models generalize across different populations. Second, they emphasized the importance of “explainable AI,” developing tools to understand which biomarkers are driving the models’ predictions and ensuring that the results are biologically meaningful and clinically trustworthy. Finally, they called for prospective testing to evaluate whether AI-guided decision-making can safely reduce the need for OFCs while maintaining diagnostic accuracy.

The team is also exploring how these models could be integrated into clinical workflows to provide clinicians with real-time risk assessments and support shared decision-making with families. Machine learning is increasingly being applied to complex medical challenges, offering the potential to personalize and improve patient care.

Carla M. Davis, MD, can be reached at [email protected].
Kyung Won Lee, PhD, can be reached at [email protected].
Mayte Suarez-Farinas, PhD, can be reached [email protected].
McKenzie Williams can be reached at [email protected].