Prof Brian D Green (Queen's University Belfast), Dr James M Brown (University of Lincoln), Professor Jayne V Woodside (Queen's University Belfast), Damian O’Kelly (Nutritics Limited).

Scientific background:

The human diet is the principal source of energy for sustaining mankind, and it plays a pivotal role in preventing chronic disease. Maintaining a sustainable and nutritionally-balanced diet is complicated by factors such as, food quality, dietary diversity, culinary practices, and differing nutrient absorption across the population. Recent decades have seen the shift of dietary patterns towards more processed and energy-dense foods, and a decline in fruit and vegetables consumption. Overall, this has contributed to global rises in diet-related non-communicable disease. Research into dietary frameworks, such as the Mediterranean diet, aim to capitalise on their comprehensive nutritional profiles and associated health benefits. Complying with such dietary framework necessitates personalised monitoring, but traditional dietary assessment tools, including 24-hour dietary recalls and food frequency questionnaires, are limited by the fact that data are self-reported, time-consuming and potentially inaccurate.

This project will investigate the use of AI technologies in nutritional science in several ways. First, it will examine if traditional food data collection can be enhanced by AI[1]. It will use machine learning tools to identify novel combinatorial biomarkers, which could be used as objective measures of a person’s diet. Finally, it will aim to demonstrate if automated and individualised dietary assessment tools based on food images could be an accurate and effective means of promoting optimal nutrition and mitigating diet-related health risks[2].

Research methodology: 

Multimodal large language models (MLLMs) will be used to emulate the workflow of nutrition experts but using image-based food analysis and subsequent nutrient calculation[3]. The approach will capitalise on the compositional reasoning of MLLMs without demanding extensive fine-tuning. Once established, the approach will be modified to: better estimate food volume (portion size) using monocular depth estimation, incorporate ‘hidden’ food ingredient data (e.g., oil), and adopt UK-centric food datasets. The AI approaches will be evaluated against a food diary from a cohort "in the wild".

Training:

The appointed PhD student gain specialised skills in:

1. Nutritional sciences, including traditional dietary assessment tools, biomarker identification, data cleaning and statistical analysis.

2. Computer sciences, including handling of large scale, real-world tabular and image data, implementation of state-of-the-art neural network architectures and training regimes, use of high-performance computing resources.

Applicants

This project welcomes applicants either from, a Computer Sciences/AI background with an interest in nutrition, or from a Nutrition/Life Sciences background with an interest in AI.