Department of Computer Science: MSc Thesis Presentation

Daniel Hopkins will present his MSc thesis "Predicting the Glycemic Impacts of Foods in Context" on Thursday 18 November at 13:00 via Zoom.

Predicting the Glycemic Impacts of Foods in Context

Author: Daniel Hopkins
Supervisor: Juho Rousu
Advisor: Markus Heinonen

Date: Thursday 18 November 2021
Time: 13:00-13:30


There is limited research on modeling the glycemic impact of real-world meals in individuals without diabetes. Research has typically focused either on isolated ingredients in controlled contexts, as in Glycemic Index (GI) measurements, or on predicting hypo- and hyperglycemia episodes in individuals with diabetes. This thesis seeks to address this gap. We use a novel dataset containing more than twenty-thousand tagged meals from nearly one-thousand individuals. We explore various models for predicting the glycemic response to a meal and discuss the performance and interpretability of each model. We explore which contextual features are useful for prediction and how they might interact.

In the modeling task, we address two questions. First, is it realistic to predict glucose responses to meals given only the meal ingredients and basic user information? And second, do these models offer any actionable insights for individuals? In particular, we use these models to explore two possible insights. First, do food ingredients exhibit interesting or unexpected interaction effects with other ingredients or contextual factors? And second, do individuals exhibit personalized responses to food ingredients, food combinations or other contextual factors. Answering the first question would lead to a better understanding of dietary science. Answering either question would lead to better guidance for individuals aiming to improve their metabolic health.

The results of this work are largely negative, which we hope will be useful for others following the same path. Predicting glucose responses is a challenging task, even with well-labeled data and exact nutritional information. In a real-world context, where data is limited, mislabeled and incomplete, it is even more so.

  • Published:
  • Updated: