2024-07-10 –, Else (1.3)
To improve our understanding of biological processes, dynamic models are built that describe the human blood glucose after a meal. However, these models are typically built based on population averages and therefore cannot capture the full scope of individual dynamics. Using SciML to incorporate machine learning techniques into these models, we aim to improve model personalisability.
Precision nutrition aims to provide personalised dietary recommendations in order to improve overall metabolic health, opposed to general dietary recommendations. By using the Eindhoven Diabetes Simulator (EDeS), a mechanistic model of human response of blood glucose and insulin to standardised meals, we aim to capture relevant and biologically interpretable parameters of current metabolic health. However, this model was built to describe population average responses and fails to describe the full extent of inter-individual variability.
To improve the model's ability to capture personalised meal responses, and to improve the relevance of individual model parameter values, the framework of neural universal differential equations was used. By incorporating a neural network into targeted components of the model, we can specifically increase the model's flexibility, without compromising strong biological bias that enables parameter estimation on limited data.
In this talk, we present results on training universal differential equations on sparsely sampled human data using physiology-informed regularisation and the recovery of personalised glucose-driven insulin production using conditional neural networks in universal differential equation models.
In this way, the hybrid model enables us to improve our understanding of individual variation in blood sugar responses to meals, and provides interpretable measures to aid in the design of personalised dietary recommendations.
I’m a PhD candidate in systems biology for metabolic disease at the Department of Biomedical Engineering at Eindhoven University of Technology. I am working on model personalisation with scientific machine learning.