Accurate Blood Glucose (BG) prediction is essential for enabling glycemic control in individuals with Type 1 Diabetes Mellitus (T1DM), particularly within Smart and Connected Health (SCH) systems that integrate Continuous Glucose Monitoring (CGM) and automated insulin delivery. The adaptability of Large Language Models (LLMs) provides a promising foundation for unified, fine-tunable forecasting models. We introduce DiabLLM, a framework based on two recent LLM-based architectures: Time-LLM, which incorporates a lightweight projection layer and alignment techniques to transform time-series data into embeddings interpretable by pre-trained LLMs, and Chronos, which employs time-series-aware tokenization and quantization to convert continuous inputs into discrete sequences for forecasting. Both models process 30-minute sequences of six historical BG values and predict 30- and 45-minute horizons. Experimental results on the OhioT1DM and D1NAMO datasets demonstrate that DiabLLM outper forms state-of-the-art baselines, including a Deep Reinforcement Learning model and an ensemble of LSTM, GRU, and WaveNet, achieving up to 27% improvement in RMSE and 37% in MAE. To enhance robustness to noisy and missing input data, a denoising autoencoder was employed for input reconstruction, yielding improved predictive performance. In addition, knowledge distillation was shown to significantly compress the model, making it a practical candidate for efficient deployment on resource-constrained edge devices without compromising accuracy.
Mahmoudi et al. (Thu,) studied this question.