Menu labeling, calories, and nutrient density: Evidence from chain restaurants

An Explainable Prediction for Dietary-Related Diseases via Language Models

Our study harnesses the power of natural language processing (NLP) to explore the relationship between dietary patterns and metabolic health outcomes among Korean adults using data from the Seventh Korea National Health and Nutrition Examination Survey (KNHANES VII). Using Latent Dirichlet Allocation (LDA) analysis, we identified three distinct dietary patterns: "Traditional and Staple", "Communal and Festive", and "Westernized and Convenience-Oriented". These patterns reflect the diversity of dietary preferences in Korea and reveal the cultural and social dimensions influencing eating habits and their potential implications for public health, particularly concerning obesity and metabolic disorders. Integrating NLP-based indices, including sentiment scores and the identified dietary patterns, into our predictive models significantly enhanced the accuracy of obesity and dyslipidemia predictions. This improvement was consistent across various machine learning techniques-XGBoost, LightGBM, and CatBoost-demonstrating the efficacy of NLP methodologies in refining disease prediction models. Our findings underscore the critical role of dietary patterns as indicators of metabolic diseases. The successful application of NLP techniques offers a novel approach to public health and nutritional epidemiology, providing a deeper understanding of the diet-disease nexus. This study contributes to the evolving field of personalized nutrition and emphasizes the potential of leveraging advanced computational tools to inform targeted nutritional interventions and public health strategies aimed at mitigating the prevalence of metabolic disorders in the Korean population.

What's On the Menu? Towards Predicting Nutritional Quality of Food Environments

Unhealthy diets are a leading cause of major chronic diseases including obesity, diabetes, cancer, and heart disease. Food environments-the physical spaces in which people access and consume food-have the potential to profoundly impact diet and related diseases. We take a step towards better understanding the nutritional quality of food environments by developing MINT: Menu Item to NutrienT model. This model utilizes under-studied data sources on recipes and generic food items, along with state-of-the-art word embedding and deep learning methods, to predict the nutrient density of never-before-seen food items using only their name as input. The model achieves an R 2 = 0.77 , a substantial improvement over comparable models. We illustrate the utility of MINT by applying it to the Los Angeles restaurant food environment, and discover close agreement between predicted and ground truth nutrient density of restaurant menu items. This model represents a significant step towards a policy toolkit needed to precisely identify and target food environments characterized by poor nutritional quality.