Personalized glucose forecasting for people with type 1 diabetes using large language models

Area of Science:

• Artificial Intelligence
• Endocrinology
• Biomedical Informatics

Background:

• Type 1 Diabetes (T1D) management necessitates precise glucose level control using exogenous insulin.
• Continuous Glucose Monitoring (CGM) has advanced diabetes care, yet glucose variability remains a challenge.
• Large Language Models (LLMs), successful in text, are unexplored for multimodal data like glucose forecasting.

Purpose of the Study:

• Evaluate LLM-based models for glucose forecasting in T1D.
• Compare prediction performance between insulin Multiple Daily Injections (MDIs) and pump users.
• Develop a personalized, adaptive glucose prediction model.

Main Methods:

• Utilized CGM data from the T1DEXI study for glucose level forecasting.
• Assessed various predictive models, including LLM-based approaches.
• Evaluated model performance using Mean Absolute Error (MAE) across 60, 90, and 120-minute prediction horizons.

Main Results:

• Personalized LLM approaches yielded superior short-term (60-90 min) glucose predictions (MAE: 15.2-17.2 for pumps, 15.7-20.2 for MDIs).
• For long-term (120 min) prediction, TIDE (MAE: 19.8 for MDIs) and Patch-TST (MAE: 18.5) showed competitive results.
• LLM models demonstrated comparable MAE to state-of-the-art methods with reduced variability.

Conclusions:

• LLM-based models offer a viable alternative for glucose forecasting in T1D.
• The personalized approach shows significant potential for improving short-term glycemic control.
• This research advances personalized prediction models to mitigate diabetes complications.