Related Experiment Video
Updated: Jun 27, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Empowering digital health management with on-device large language models for glucose prediction: a model development
Taiyu Zhu1, Joanna Howson2, Alejo Nevado-Holgado3
1Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford, OX3 7JX, UK; Department of Biostatistics & Health Informatics, Institute of Psychiatry, Psychology & Neuroscience, King's College London, 16 De Crespigny Park, London, SE5 8AB, UK.
Background:
Long-term management of chronic diseases such as diabetes is increasingly based on wearable technologies, particularly continuous glucose monitoring (CGM), integrated with smartphone-based digital health systems. When combined with artificial intelligence, especially deep learning, these systems offer highly personalised decision support, including glucose prediction. Although large language models (LLMs) have demonstrated strong performance across various healthcare tasks, their integration into day-to-day digital health remains limited, primarily due to privacy concerns associated with transmitting sensitive data to remote servers. Recent advances in lightweight LLMs create new opportunities for secure and local deployment.
Methods:
In this study, we first evaluated the zero-shot glucose prediction performance of eight pretrained lightweight LLMs across multiple model families. None achieved clinically viable outputs, highlighting the need for domain-specific adaptation. To address this, we propose GluLLM, a multimodal adaptor-based framework that enhances pretrained LLMs for on-device glucose forecasting. GluLLM integrates CGM data, daily activity logs, and electronic health records using customised encoder and decoder modules while preserving the foundational capabilities of pretrained LLMs. We trained and evaluated GluLLM on the REPLACE-BG dataset, which includes 226 individuals with type 1 diabetes, and validated it on an external cohort comprising 207 individuals with type 2 diabetes or without diabetes.
Findings:
Compared with 15 state-of-the-art deep learning baselines for time-series prediction, GluLLM (LLaMA 3.2 1B backbone) demonstrated superior performance, with significantly lower 30-min root mean square error than the strongest baseline (Crossformer) on REPLACE-BG and Móstoles (20.6 ± 3.5 and 9.6 ± 2.9 mg/dL; p < 0.001), and improved hypoglycaemia prediction (glucose <70 mg/dL; AUROC: 0.79 and 0.84; AUPRC: 0.55 and 0.60), respectively. Furthermore, deployment of GluLLM on two smartphone platforms demonstrated feasible computational requirements, with acceptable CPU and memory usage and low inference latency.
Interpretation:
GluLLM demonstrates that LLMs can support the next generation of smartphone-based digital health systems, delivering real-time, privacy-preserving clinical decision support.
Funding:
Novo Nordisk Postdoctoral Fellowship run in partnership with the University of Oxford.