Personalized sleep advice from wearable data using small language models
Toward Personalized Sleep Guidance from Wearable Data Using Language Models
Computation and LanguageArtificial Intelligence
Summary
Sleep trackers collect a lot of data, but giving clear, personal advice from that data is hard. The authors created a two-step method where big language models first turn sleep data into useful tips without needing expert labels. Then, they train smaller, faster models to give quick and private sleep advice. Their approach works better than existing general or medical models, and people find the advice helpful.
What this means in practice
- •For wearable device developers: Integrate lightweight personalized sleep advice models directly on devices to enhance user privacy and accessibility.$Commercial implications: This paper enables on-device sleep guidance tools that protect user privacy and reduce dependency on cloud processing.
- •For health app developers: Add customized sleep summaries and recommendations based on raw wearable data without costly expert annotation.
Authors
Yusheng Tan, Running Zhao, Sofia Angel, Ninghui Hao, Ash Arian, Nikita N. Dulin, Jay Lin, Ou Zhu, Faiza Shaik, Xinxing Yang, Bonnie W. Leung, Katie Roster, Arlene Ruiz de Luzuriaga, Kenneth Lee, Alejandra Lastra, Habibul Ahsan, Guihong Wan
Abstract
Sleep monitoring using wearable data has shown promise for personal health, yet large language model (LLM)-based summarization and question answering remain insufficient for personalized sleep guidance. Training specialized models, however, often requires costly expert annotation. Moreover, privacy and accessibility concerns motivate lightweight, local deployment for end users. We present a two-stage framework to address these challenges. Specifically, in Stage~1, a multi-agent LLM pipeline reasons structured sleep guidance from unannotated wearable records, enabling scalable dataset construction. Stage~2 distills guidance reasoning trajectories into small language models (SLMs) through supervised fine-tuning and integrates a training-free Best-of-$N$ selection strategy to enhance inference. Experimental results demonstrate our method outperforms commercial general and medical LLMs and open-source models. Human evaluation further supports the quality of the generated guidance and the feasibility of personalized sleep guidance with SLMs.