Multi-Objective Optimisation for Personalised Nutrition: Constraint-Based Meal Planning, Explainable AI, and Adaptive Feedback

About the project

About the Project Healthcare platforms for personalised nutrition have predominantly focused on invasive biomarker prediction, such as continuous glucose monitoring or microbiome profiling, to generate dietary recommendations. This project instead asks how non-invasive, interpretable optimisation techniques can address the dietary needs of individuals managing clinical conditions such as Celiac disease, Type 2 diabetes, or medication-specific nutritional restrictions. The central research question is whether constraint-based multi-objective optimisation can generate meal plans that balance nutritional adequacy, budget constraints, preparation effort, and clinical safety without relying on invasive biological sampling. The research is designed to avoid plain fitness-app territory by focusing on medically relevant dietary constraints, implicit preference learning, and explainable decision-making. Prior work in operational research has applied multi-objective evolutionary algorithms to menu planning for institutional settings, but these have typically treated cost and nutritional adequacy as static objectives and have not incorporated individualised health constraints, dynamic user preferences, or real-world supply variability. This project extends that foundation. Research Objectives : 1. Develop a mathematical framework to balance competing objectives: macro/micro-nutrient adequacy, cost minimisation, and user effort (preparation time and ingredient variety). Prior work in operational research has applied multi-objective evolutionary algorithms to menu planning for institutional settings, but these have typically treated cost and nutritional adequacy as static objectives and have not incorporated individualised health constraints or preference dynamics. This research asks: how can these formulations be extended to account for individualised health constraints and personal preference dynamics, and what algorithmic structures are needed to maintain optimality? 2. Create a formal knowledge-base encoding interactions between common medications, genetic predispositions, family history, and nutritional intake to ensure meal plans are clinically safe and condition-specific. Prior efforts include drug-nutrient interaction databases and clinical decision support systems for dietetics, but these have not been integrated as formal constraints within optimisation engines. This research asks: how can these existing knowledge resources be formalised as operational constraints, and what inference mechanisms are needed to resolve conflicts when multiple clinical constraints compete? 3. Investigate Inverse Reinforcement Learning (IRL) or preference-learning algorithms that infer a user's true dietary tastes from implicit signals, such as plan modifications, rejections, or ingredient substitutions, rather than relying on burdensome static surveys. Prior work in recommender systems has explored implicit feedback for food preference modelling, but Inverse Reinforcement Learning remains underexplored in the nutritional optimisation domain. This research asks: can IRL recover reward functions that represent latent dietary preferences from sparse interaction logs, and how do these learned preferences interact with the clinical and budget constraints above? 4. Investigate how low-frequency IoT signals from wearables, such as step count, sleep quality, or heart rate variability, can be fused into the optimisation model as dynamic constraints, enabling plan adjustment without invasive biomarker sampling. Prior work has linked physical activity data to caloric recommendation systems, but the integration of wearables into formal constraint optimisation remains nascent. This research asks: what temporal abstraction and uncertainty quantification methods are needed to convert noisy wearable signals into reliable optimisation constraints, and how does this affect plan stability over time? 5. Develop a simulated retail environment to evaluate how the optimisation engine maintains nutritional integrity under supply variability. Rather than building commercial API integrations, the student will construct a controlled simulation that models inventory fluctuations, seasonal price variation, and product substitution. This research asks: under what conditions do supply-side perturbations degrade the nutritional quality of generated plans, and what robustness mechanisms can be designed to preserve constraint satisfaction? Methodology : The student will employ multi-objective evolutionary algorithms or constraint satisfaction programming to navigate the high-dimensional search space of meal planning, drawing on existing libraries such as OR-Tools and DEAP. Evaluation will use public nutritional databases, including the McCance and Widdowson dataset, and the simulated retail environment described above. Controlled user studies will be conducted to gather implicit preference signals, but will be designed as experiments rather than longitudinal deployments. The methodological approach spans optimisation theory, knowledge representation, machine learning, and human-computer interaction. Impact : The research will contribute to the University's Knowledge Exchange and Research Excellence Framework profiles through supervised PhD output and algorithmic innovation in health informatics. The work is aligned with the KU Technology Transfer Office invention disclosure pathway, ensuring IP clarity and providing a clear route to Seedcorn and subsequent ICURe Explore funding. The candidate will be expected to publish in venues such as the ACM Conference on Recommender Systems or the International Journal of Medical Informatics. Desirable Skills : Strong programming skills; familiarity with optimisation libraries such as OR-Tools or DEAP; interest in Health Informatics, Digital Health, or Operational Research. Passion for healthcare and personalisation is essential. Experience with wearable or IoT data, and an understanding of API integration, would be advantageous.

Research areas