DiabLLM: An LLM-Based Framework for Blood Glucose Prediction in Type 1 Diabetes

Area of Science:

• Artificial Intelligence in Healthcare
• Biomedical Informatics
• Time Series Forecasting

Background:

• Accurate blood glucose (BG) prediction is critical for glycemic control in Type 1 Diabetes Mellitus (T1DM).
• Smart and Connected Health (SCH) systems require reliable BG forecasting for integrated Continuous Glucose Monitoring (CGM) and automated insulin delivery.
• Large Language Models (LLMs) offer adaptable architectures for developing unified, fine-tunable forecasting models.

Purpose of the Study:

• To introduce DiabLLM, a framework leveraging LLM-based architectures for improved BG prediction in T1DM.
• To evaluate the performance of DiabLLM against state-of-the-art models on established datasets.
• To assess methods for enhancing model robustness and enabling efficient deployment on edge devices.

Main Methods:

• DiabLLM framework integrates Time-LLM and Chronos architectures for time-series data processing.
• Models transform historical BG data into interpretable embeddings or discrete sequences for LLM forecasting.
• A denoising autoencoder was used for input data reconstruction, and knowledge distillation for model compression.

Main Results:

• DiabLLM outperformed state-of-the-art baselines on OhioT1DM and D1NAMO datasets, with up to 27% RMSE and 37% MAE improvement.
• Input reconstruction using a denoising autoencoder enhanced predictive performance, particularly with noisy or missing data.
• Knowledge distillation achieved significant model compression, enabling practical deployment on resource-constrained edge devices without accuracy loss.

Conclusions:

• DiabLLM provides a robust and accurate framework for BG prediction, advancing T1DM management within SCH systems.
• The framework demonstrates the potential of LLMs for complex biomedical time-series forecasting.
• DiabLLM's efficiency and robustness make it suitable for real-world, on-device applications in diabetes care.