AI Model Shows Promise for Predicting Peanut Allergy, But More Testing Is Needed

Artificial Intelligence is showing promise as a potential tool for one of the most difficult questions in peanut allergy diagnosis: does a positive allergy test indicate a genuine clinical allergy, or is the patient simply sensitised to peanut without reacting when they eat it?

A new study from researchers at Charité–Universitätsmedizin Berlin has developed explainable machine-learning models that combine clinical and immunological information to predict the outcome of an oral food challenge. The strongest model achieved 96% accuracy during cross-validation and 86% accuracy when tested against an independent cohort.

The findings could eventually help clinicians identify which peanut-sensitised patients are more likely to have a genuine allergy and which may be able to tolerate peanut. That could have practical implications because oral food challenges remain the gold standard for confirming food allergy, but they are resource-intensive and can trigger potentially severe allergic reactions.

Importantly, however, the study is a proof of concept rather than a ready-to-use clinical AI system. The researchers say larger, prospective and multicentre studies will be necessary before these models can be incorporated into routine allergy care.

Why peanut sensitisation does not always mean peanut allergy

One of the central challenges in allergy medicine is that sensitisation and clinical allergy are not the same thing.

A person may have detectable antibodies against peanut allergens without experiencing symptoms after eating peanut. This is known as sensitisation. A person with a genuine peanut allergy, on the other hand, develops a clinical reaction when exposed to the allergen.

This distinction matters because commonly used allergy tests can identify evidence of an immune response, but they do not always establish whether eating the food will actually cause symptoms.

That creates uncertainty for both clinicians and patients. A patient who is incorrectly labelled as allergic may unnecessarily avoid peanut, while failing to identify a genuine allergy could expose the patient to a potentially dangerous reaction.

The oral food challenge can help resolve this uncertainty because it directly observes the patient’s response to controlled exposure.

Why oral food challenges remain the diagnostic benchmark

An oral food challenge (OFC) involves giving a patient gradually increasing amounts of a suspected allergen under medical supervision while monitoring for a reaction.

Because it measures the patient’s actual response to food exposure, the OFC remains the reference standard for confirming food allergy in appropriate clinical circumstances.

However, the procedure has obvious limitations.

It requires specialist supervision, appropriate medical facilities and sufficient time for careful observation. More importantly, a patient who is genuinely allergic may experience a serious reaction during the challenge.

This creates a difficult balance for allergists. An OFC can provide valuable diagnostic certainty, but performing the test is not risk-free.

A reliable prediction model could potentially help clinicians decide when the information obtained from an OFC is worth the risk and resources involved.

How the new AI study worked

The researchers analysed data from 96 peanut-sensitised participants aged between one and 56 years.

Of the participants, 74 had confirmed peanut allergy, while 22 were sensitised but tolerant to peanut.

The researchers developed machine-learning models designed to predict whether participants would have an allergic reaction during an oral food challenge.

Three approaches were evaluated:

• Clinical data combined with basophil activation test data
• Clinical data alone
• Basophil activation test data alone

The researchers also used explainable AI techniques to identify which measurements had the greatest influence on the model’s predictions.

This is particularly relevant in healthcare because clinicians need more than an unexplained numerical prediction. Understanding which clinical and biological characteristics influence an algorithm’s decision can make AI tools easier to evaluate and potentially integrate into clinical workflows.

The combined model achieved 96% accuracy

The model combining clinical and basophil activation test data performed best during cross-validation, achieving 96% accuracy in predicting oral food challenge outcomes.

However, one of the study’s most interesting findings was that the clinical-only model performed almost as well.

The result raises an important practical question. If clinical information alone can produce predictions close to those generated by a more complex model, the simplest effective model may ultimately be easier to implement in everyday allergy clinics.

Additional laboratory testing may still provide value, but AI development should not automatically equate more data with better clinical decision-making.

Ara h 2 was one of the strongest predictors

The explainable AI analysis offered insight into which measurements were driving the predictions.

IgE antibodies against Ara h 2, one of the major peanut allergens, were particularly influential. Higher Ara h 2-specific IgE levels were associated with an allergic response during the oral food challenge.

Peanut skin-test results also contributed significantly. Larger reactions during skin testing were associated with an increased likelihood of an allergic response.

The models also incorporated information from the basophil activation test, which evaluates how basophils respond when exposed to peanut allergens.

These findings are important because they show that the AI system was not simply making a prediction from an unexplained collection of numbers. The researchers were able to identify biologically meaningful variables associated with the prediction.

Why explainable AI matters in allergy medicine

Artificial intelligence is increasingly being investigated across medicine, but healthcare applications face a particular challenge: clinicians need to understand how a system reaches its conclusions.

A so-called black-box algorithm may produce an accurate prediction, but if doctors cannot understand the factors influencing that prediction, it can be difficult to assess whether the result is clinically sensible.

Explainable AI attempts to address this problem by showing which variables contribute most strongly to an algorithm’s output.

In the peanut allergy study, the influence of Ara h 2-specific IgE and skin-test measurements gives clinicians a clearer picture of why the model considers a patient more likely to react.

That does not make the algorithm automatically correct. It does, however, provide a more transparent foundation for evaluating its predictions.

The independent cohort is an important test

The researchers did not stop after evaluating their models within the original study population. They also tested the models using data from an independent cohort associated with the LEAP study.

The combined model correctly classified 86% of participants in this separate dataset, while the clinical-only model achieved 85% accuracy.

The difference between the original cross-validation performance and the independent-cohort performance is an important lesson in medical AI.

A model can perform extremely well when tested using data that resemble the information used during its development. Performance may change when the model encounters patients from another study or population.

That is why external validation is essential before an AI system can be trusted in routine clinical practice.

The biggest limitation: predicting severity was much harder

The AI models were substantially better at answering a relatively straightforward question: is this patient likely to have an allergic reaction?

They were less successful when researchers attempted to predict more detailed outcomes, including maximum tolerated peanut dose and reaction severity.

This distinction is clinically important.

Predicting whether an allergy is likely to be present is different from predicting exactly how much allergen a person can tolerate or how severe their reaction will be.

The latter requires a much greater degree of precision because an incorrect prediction could have direct consequences for patient safety.

The weaker performance of the regression models therefore reinforces the need for caution. AI may eventually help with diagnostic classification without necessarily being capable of accurately predicting the full clinical course of an allergic reaction.

Could AI reduce the number of oral food challenges?

Potentially, but that is not yet established.

The most realistic future role for this technology would be as a clinical decision-support tool.

An allergist could potentially use an AI prediction alongside the patient’s medical history, allergy-test results and clinical assessment when deciding whether an oral food challenge is appropriate.

For example, a patient whose clinical and laboratory profile strongly suggests allergy might be identified as less suitable for a diagnostic challenge, while a patient with conflicting evidence might be identified as someone for whom a supervised challenge could provide valuable information.

However, the AI system should not be viewed as independently deciding whether a patient can safely eat peanut.

That distinction is essential. A prediction model can support clinical reasoning, but it cannot remove the need for professional judgment and appropriate patient safety measures.

The surprising finding: more testing did not produce a dramatic improvement

One of the study’s most interesting insights is the relatively small difference between the combined model and the clinical-only model.

The combined model achieved 96% accuracy, compared with 95% for the clinical-only model during cross-validation.

This suggests that the clinical information already contained substantial predictive value.

From a healthcare-system perspective, this could eventually prove useful. If a reliable model can achieve strong performance using information already collected during routine allergy assessment, implementation could be simpler than systems requiring extensive additional laboratory testing.

At the same time, the study was not large enough to establish that clinical data will consistently outperform or match laboratory-enhanced approaches across different populations.

That question will require further research.

What this study does not prove

The findings are promising, but they should not be interpreted as evidence that AI is ready to replace oral food challenges or allergists.

• The study included only 96 peanut-sensitised participants.
• The participants represented a specific clinical population rather than everyone being evaluated for possible peanut allergy.
• The strongest 96% performance came from cross-validation within the study data.
• Accuracy fell to 86% when the combined model was tested against an independent cohort.
• The models were less successful at predicting reaction severity and maximum tolerated dose.
• The researchers describe the work as a proof of concept.
• Larger prospective and multicentre validation studies are still required.

These limitations are particularly important in allergy medicine, where an incorrect assessment can expose a patient to a potentially serious reaction.

What would need to happen before clinical adoption?

Before such a model could become part of routine allergy care, researchers would need to demonstrate that it works reliably across larger and more diverse populations.

Future studies should include patients from multiple centres and healthcare systems, with different demographic characteristics and clinical profiles.

Researchers would also need to determine whether using the AI model actually improves patient care.

High accuracy alone is not enough. A successful clinical tool should ideally reduce unnecessary procedures, improve diagnostic confidence or make the pathway to an accurate diagnosis more efficient without increasing risk.

Another important question will be how clinicians should respond when the AI prediction conflicts with their own assessment.

For a high-stakes diagnosis such as peanut allergy, the safest model is likely to keep the clinician in control while using AI as an additional source of evidence.

Why this could matter for patients

A reliable diagnostic prediction system could have benefits beyond reducing the number of oral food challenges.

Patients incorrectly labelled as peanut-allergic may avoid peanut unnecessarily for years. Such restrictions can affect everyday eating, social situations and quality of life.

At the other end of the spectrum, accurately identifying genuine peanut allergy allows patients and families to receive appropriate counselling about avoidance, emergency preparedness and ongoing allergy management.

AI could potentially help clinicians distinguish these groups more efficiently, particularly in cases where conventional test results are difficult to interpret.

The bigger future: personalised allergy diagnosis

The study points toward a broader shift in allergy medicine from interpreting individual test results toward integrating multiple sources of patient information.

Instead of asking whether one test is positive or negative, future diagnostic systems could combine clinical history, skin-test results, allergen-specific IgE measurements and functional immune testing to estimate the probability of a genuine clinical allergy.

Machine learning is particularly suited to this type of pattern recognition because it can evaluate relationships among multiple variables simultaneously.

But the ultimate goal should not be to make allergy diagnosis more complicated. It should be to make it more accurate and more personalised while maintaining patient safety.

Conclusion

The new research provides encouraging evidence that explainable artificial intelligence could help distinguish peanut allergy from peanut sensitisation.

The strongest model, which combined clinical and basophil activation test data, achieved 96% accuracy during cross-validation and 86% accuracy in an independent cohort. Clinical data alone performed nearly as well in the initial analysis, while BAT-only modelling was less accurate.

The findings also highlight the current limits of the technology. Predicting whether a patient will react was considerably easier than predicting reaction severity or maximum tolerated dose, and the researchers emphasise the need for larger prospective validation.

For now, oral food challenges remain an important diagnostic tool and AI should be viewed as a potential future aid rather than a replacement for specialist assessment.

If future studies confirm these results across diverse populations, explainable AI could eventually help allergists identify patients who are most likely to benefit from an oral food challenge, reduce unnecessary diagnostic procedures and improve confidence when interpreting complex allergy-test results.

The most promising future for AI in peanut allergy may not be replacing the allergist, but helping the allergist make a more informed decision about which patients truly need the most demanding diagnostic test.