Integrating LLM-Based Nutrient Estimation from Chinese Diet Logs for Better Continuous Glucose Forecasting

Accurate short-term forecasting of blood glucose is essential for safe diabetes self-management, yet existing forecasters seldom leverage nutrients estimated from free-text meal logs. This estimating nutrient is especially challenging for Chinese meals due to mixed dishes and diverse preparation styles. In this study, we investigate whether large language model (LLM)-based nutrient estimation from Chinese diet logs can improve dynamic glucose forecasting. Five Chinese-capable LLMs for Chinese texts (GPT5-mini, DeepSeekv3.1, Llama3.170B, GLM4.5, Qwen3-235B) are prompted to estimate macronutrients and weighted GI for meals. We also compare two pipelines for nutrition estimation: database-first with LLM fallback and LLM-only with retrieval-augmented generation (RAG). We propose a seasonal-patch Transformer architecture designed for multi-source feature fusion, integrating dynamic sequences, static patient data, diet word embeddings, and the LLM-estimated nutritional profiles. Experiments on 112 realworld Chinese patients (ShanghaiT1DM $\mathrm{n}=12$, ShanghaiT2DM $\mathbf{n}=\mathbf{1 0 0}$) showed that, versus a no-LLM baseline, LLMaugmented models reduced 3-h postprandial RMSE by up to 1.2 $\mathbf{m g} / \mathbf{d L}(\mathbf{3 6. 0 6} \rightarrow \mathbf{3 4. 8 6 ~ m g} / \mathbf{d L})$ and achieved the best full-period RMSE of $26.66 \text{mg} / \text{dL}$, comparable to a recent large pretraining forecasting model. Forecast gain correlates strongest with carbohydrate $(\rho=0.94)$ and fat $(\rho=0.77)$ estimation accuracy. These results demonstrate that LLM-based nutritional feature engineering provides a critical pathway toward more precise, personalized CGM forecasting and the realization of clinical digital twins.