Related Experiment Video
Updated: Jun 6, 2025
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
Translational Algorithms for Technological Dietary Quality Assessment Integrating Nutrimetabolic Data with Machine
Víctor de la O1,2, Edwin Fernández-Cruz1,2, Pilar Matía Matin3,4
1Cardiometabolic Nutrition Group, Precision Nutrition Program, Research Institute on Food and Health Sciences IMDEA Food, Consejo Superior de Investigaciones Científicas-Universidad Autónoma de Madrid (CSIC-UAM), 28049 Madrid, Spain.
Abstract:
Recent advances in machine learning technologies and omics methodologies are revolutionizing dietary assessment by integrating phenotypical, clinical, and metabolic biomarkers, which are crucial for personalized precision nutrition. This investigation aims to evaluate the feasibility and efficacy of artificial intelligence tools, particularly machine learning (ML) methods, in analyzing these biomarkers to characterize food and nutrient intake and to predict dietary patterns.
Methods:
We analyzed data from 138 subjects from the European Dietary Deal project through comprehensive examinations, lifestyle questionnaires, and fasting blood samples. Clustering was based on 72 h dietary recall, considering sex, age, and BMI. Exploratory factor analysis (EFA) assigned nomenclature to clusters based on food consumption patterns and nutritional indices from food frequency questionnaires. Elastic net regression identified biomarkers linked to these patterns, helping construct algorithms.
Results:
Clustering and EFA identified two dietary patterns linked to biochemical markers, distinguishing pro-Mediterranean (pro-MP) and pro-Western (pro-WP) patterns. Analysis revealed differences between pro-MP and pro-WP clusters, such as vegetables, pulses, cereals, drinks, meats, dairy, fish, and sweets. Markers related to lipid metabolism, liver function, blood coagulation, and metabolic factors were pivotal in discriminating clusters. Three computational algorithms were created to predict the probabilities of being classified into the pro-WP pattern. The first is the main algorithm, followed by a supervised algorithm, which is a simplified version of the main model that focuses on clinically feasible biochemical parameters and practical scientific criteria, demonstrating good predictive capabilities (ROC curve = 0.91, precision-recall curve = 0.80). Lastly, a reduced biochemical-based algorithm is presented, derived from the supervised algorithm.
Conclusions:
This study highlights the potential of biochemical markers in predicting nutritional patterns and the development of algorithms for classifying dietary clusters, advancing dietary intake assessment technologies.
