Profile-Aided Distillation Framework for Personalized Sleep Analysis with Compact Models Using LLM-Guided Synthetic Data

Huimin Zheng^1*Xingxing Ai²Xueyan Liu³Xiaofen Xing¹Xiangmin Xu¹

• ¹South China University of Technology, Guangzhou, China
• ²Jinan University, Guangzhou, China
• ³The First Affiliated Hospital of Jinan University, Guangzhou, China

Enabling personalized sleep analysis and interaction directly on edge devices is crucial for providing real-time health insights and tailored guidance, yet remains challenging due to data scarcity and computational constraints. In this study, we propose a novel framework for personalized sleep analysis on edge devices, addressing two critical obstacles: the scarcity of publicly available physiological datasets and the limited capacity of compact models. To mitigate data scarcity, we introduce a Physiologically-Constrained Adaptive Hierarchical Copula approach, which leverages large language model-guided optimization to synthesize diverse and realistic physiological data. To enhance the limited personalized inference capabilities of compact models, we develop a Profile-Aided distillation of expert inference with MoE LoRA, which integrates user-specific information to enhance the performance of edge-deployed models. This strategy enables efficient, individualized inference and decision-making, while overcoming the limitations inherent in standard LoRA fine-tuning. Extensive experiments on both public and in-house datasets demonstrate that our distilled models achieve performance on par with state-of-the-art LLMs, yet operate efficiently within the constraints of edge devices. Overall, this work presents a promising solution for enabling personalized sleep analysis and user interaction in resource-constrained environments.