Scientific Machine Learning for Glucose Management in Diabetes

Overview

We are interested in developing data-driven approaches for glucose prediction and control in diabetes management by leveraging state-of-the-art machine learning architectures such as large language models (LLMs), transformers, and State Space Models (SSMs). These models are particularly effective in capturing long-term dependencies and complex temporal dynamics from multivariate time-series data collected through wearable devices, including continuous glucose monitors (CGMs), insulin pumps, and physical activity sensors. By integrating these diverse physiological signals, we aim to construct patient-specific models that can deliver accurate and timely glucose forecasts. Furthermore, we incorporate scientific machine learning techniques to embed domain knowledge and physiological constraints into the learning process, enhancing interpretability, generalization, and reliability. These intelligent systems will ultimately support real-time decision-making, enabling dynamic insulin dosing and adaptive therapeutic recommendations tailored to individual patient needs.