The First-of-its-kind Study Shows AI Tool Can Improve Best Practices in Managing Nut Allergies

A powerful, updatable augment for dietitians, nutritionists and allergists

By Markus Stripf, Co-CEO & Founder Spoon Guru

For up to 1 in 10 Americans who live with food allergies and their families, only daily vigilance keeps mealtime a safe pleasure.  Food shopping at retail or online can be an exhaustive effort to filter out risky products every time – because even familiar products may change their ingredients, and new items constantly come to market. 

Retailers know they’ll truly differentiate if they eliminate this customer stress, anxiety and fear, and save shoppers time.  Store-based and corporate dietitians already make a difference through their insights and recommendations.  

Even richer dialogue could occur face-to-face between these accessible health professionals and customers if they also had a robust, accurate Artificial Intelligence tool with Machine Learning capabilities that unblinkingly tracks and updates product and recipe information across the marketplace.  Such a tool, designed with input from nutritionists and allergists, would enable practical, personalized, tailored health and lifestyle recommendations based on each person’s condition, preferences and behavior – and lead to a better quality of life.  

To test this concept, Spoon Guru conducted a study to validate whether Machine Learning could perform as accurately as qualified health professionals in recommending foods suitable for people with a nut allergy.

The results of the study strongly proved that AI can be used in conjunction with healthcare professionals to help patients with nut allergy find the correct products.   

As Caroline Bovey, Honorary Chairman of British Dietetic Association puts it, ‘this study demonstrates the supportive role that evolving technologies like AI can have on health care. Dietitians already use assistive technology which significantly aids their practice and, in this case, clearly it can help patients who need to adapt their diets to access the broadest range of choices in a safe way.” -

Research Methodology  

Spoon Guru randomly sampled 2,000 food products from a database of 96,141.  Three Registered Dietitians who regularly consult food-allergy patients independently assessed each product’s information and then reached a unanimous consensus on each item’s suitability (yes or no).  

The Spoon Guru Machine Learning Model (SGML) compared against this benchmark of established ground truth.  Five more Registered Dietitians (all British Dietetic Association members who consult food-allergy patients) and three Clinical Allergists also independently compared against this benchmark.

Both healthcare professionals (HCPs) and the SGML scored on the precision and accuracy of their product assessments. Accuracy measures the proportion of responses correctly identifying a product as suitable or not suitable for those with a nut allergy.  Precision reflects the correct detection of products unsuitable for people with a nut allergy, part of correctly classifying products among four categories:

1. Product correctly defined as ‘not suitable.’
2. Product correctly defined as ‘suitable.’
3. Products incorrectly defined as ‘suitable – a false positive.’
4. Products incorrectly defined as ‘not suitable – a false negative.’  

There were no food labels to identify each item, instead health professionals saw only a spreadsheet of product names, ingredients lists and all on-pack statements. 

Study Findings Show Machine Learning Would Be a Highly Accurate Food-Assessing Complement to Healthcare Professionals Advising Customers With a Nut Allergy

The data above demonstrate that the SGML Artificial Intelligence tool can robustly enhance best practices of healthcare professionals counseling customers on their nut allergies.  Findings show AI accurately assesses extensive product information, accurately applies complex search terminology, and provides lists of suitable foods.

In the study, SGML was the most precise and made the fewest errors compared to healthcare professionals. Its measures were significantly better than six out of the eight HCPs (p<0.01), and considerably better than the second most accurate HCP (p<0.05).

Specifically, SGML was 99.8% precise with 15 errors, while practitioners were 90.5% precise on average with 183.6 errors.  The dietitians and allergists ranged in precision from 82.9% to 98.8%.

Also, SGML was 99.3% accurate, while practitioners were 90.8% accurate on average.  The dietitians and allergists ranged in accuracy from 83.1% to 99.0%.

The study doesn’t suggest that AI is superior to human performance for several reasons:  It is a small study.  The monotonous, repetitive nature of reviewing detailed spreadsheets likely led to practitioners’ fatigue and performance error and variation.  They typically wouldn’t be called on to review this volume of information in a compressed timeframe.  Also food labels under European Union regulation (study conducted in England) must embolden allergens within the ingredients list; the spreadsheet didn’t embolden allergens.

Notably, most errors made by practitioners were ‘false positives,’ where said specific foods were safe to eat, but they weren’t.  Conversely, most errors made by SGML were ‘false negatives,’ where it said some suitable foods were unsuitable, which indicates a risk-averse approach in the AI tool. 

The Practical Value of an AI Tool

An AI tool makes it possible to assess products for suitability at rates beyond our usual human capacity or availability.  For example, the SGML tool in this study can accurately evaluate 3,000 items of food per hour.  

Dietitians understand that such tools streamline their processes, so they have more time for dietary counsel, negotiate patient goals for dietary and lifestyle change, and help tailor solutions to specific patient needs.1

Findings in this study are similar to other healthcare research that demonstrate AI systems can achieve performance on par with experts.2, 3

Who Might AI Help the Most?

The latest epidemiological data suggest that of the up to 1 in 10 people affected by food allergies, those living in industrialized areas are disproportionately affected, with allergies more common in children than adults.  The foods driving the most serious burden are peanut, tree nut, fish, shellfish, egg, milk, wheat, soy and seeds.4

Proven AI tools integrated into apps used by healthcare professionals can empower them to assess more foods faster, help keep people safer, and deeply imprint their store’s wellness image all at once.

1. Chen J, Lieffers J, Bauman A, Hanning R, Allman-Farinelli M (2017 a) Designing health apps to support dietetic professional practice and their patients: qualitative results from an international survey. JMIR UHealth 5(3): e40. 
2. 2  Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S (2017) Dermatologist-level classification of skin cancer with deep neural networks. Nature 524, pages 115-118. 
3. 3  Burlina P, Pacheco KD, Joshi N, Freund DE, Bressler NM (2017) Comparing humans and deep learning performance for grading AMD: A study in using universal deep features and transfer learning for automated AMD analysis. Computers in Biology and Medicine, volume 82, pages 80-86. 
4. 4  Sicherer SH, Sampson HA (2018) Food allergy: A review and update on epidemiology, pathogenesis, diagnosis, prevention and management. Journal of Allergy and Clinical Immunology, Vol 141 Issue 1 pages 41-58.