GluLLM: Empowering digital health management with on-device large language models for glucose prediction

We propose GluLLM, a multimodal adapter-based framework that enhances pretrained LLMs for on-device glucose forecasting. GluLLM integrates CGM data, daily activity logs, and electronic health records using customized encoder and decoder modules while preserving the foundational capabilities of pretrained LLMs.

Papar Information

• Authors: Taiyu Zhu, Joanna Howson, Alejo Nevado-Holgado
• Affiliations: University of Oxford, Novo Nordisk Research Centre Oxford
• Preprint: TBA

Dataset Preparation

Dataset	Access Link
REPLACE-BG	Access from JCHR
Móstoles	Access from PLoS ONE

Usage

To train and test the model, run:

bash run.sh

Quick Start

python demo.py \
--mn llama1 \
--cache_dir your_path_to_cache_of_LLM_weights

We provide a fully self-contained synthetic demo (demo.py) that:

• Generates synthetic CGM (continuous glucose monitoring) time series for configurable numbers of virtual subjects across train / val / test splits.
• Simulates realistic diurnal glucose patterns, meal-driven spikes, random physiological variation, and bolus insulin events.
• Pre-computes Time-Dependent Information (TDI) embeddings via the LLM backbone for every patch across all sliding windows.
• Runs the full train → validate → test loop with autoregressive inference.

Directory Hierarchy

|—— .gitignore
|—— args_generator.py
|—— demo.py
|—— layers
|    |—— pjn.py
|—— main.py
|—— models
|    |—— GluLLM.py
|    |—— model_info.py
|    |—— TDI.py
|—— run.sh
|—— utils
|    |—— metrics.py
|    |—— timefeatures.py
|    |—— tools.py

Acknowledgments & References

This work was inspired by the folloing papers

• AutoTimes
• Time-LLM

We extend our gratitude to the following GitHub repositories for their valuable code and contributions:

• TSlib

License

BSD 3-Clause License

Copyright (c) 2026, University of Oxford and Novo Nordisk A/S. All rights reserved.

Citing

Please use the following BibTeX entry.

TBA