A Machine Learning-based Dietary Recommendation System for Allergy and Nutritional Risk Management

Abstract

This paper introduces Allergist, a dietary recommendation system based on machine learning for people suffering from food allergies and associated nutritional risks. As many as 1-10% of the population are estimated to have a food allergy, while navigation of everyday eating choices continues to be challenging due to ambiguous food labelling and a constant risk of cross-contamination. There are many mobile health applications in existence that, as reviewed, do not provide verified information, personalization or real-time assistance. The system proposed in this paper aims to tackle these shortcomings with a single web platform that integrates a multi-model allergen classifier, an autoencoder-based recommender, and a conversational assistant. Three classifiers (Random Forest, Gradient Boosting, and a Neural Network) were trained and evaluated on a curated collection of 54,697 recipes, each described using 883 ingredients and 54 allergen classes. Gradient Boosting had the highest micro-averaged F1 score of 100.0 percent, followed by the Neural Network with 99.5 percent, and then Random Forest at 91.2 percent. Each recipe was embedded in a 64-dimensional latent space using the autoencoder, and the adoption of cosine similarity over these results led to personalized and allergen-safe suggestions. The initial evaluation with 10 users resulted in positive scores for usability and perceived safety. The findings of this report indicate that an integrated multi-model approach could enhance the trustworthiness and availability of digital allergy support, thereby aligning with Sustainable Development Goal 3.