CGM-Agent: Privacy-Preserving AI Framework for Diabetes Data Analysis Using LLMs

A new privacy-preserving framework called CGM-Agent enables question answering over continuous glucose monitor data using large language models while keeping sensitive health information on users' devices. Continuous glucose monitors collect detailed personal health information that could enhance diabetes self-management, but existing patient platforms provide only static summaries without supporting interactive queries. The framework addresses privacy and accuracy concerns by using LLMs solely as reasoning engines that select analytical functions, with all computation occurring locally so personal data never leaves the device. For evaluation, researchers created a benchmark of 4,180 questions combining parameterized templates with real user queries and ground truth from deterministic program execution. Six leading LLMs were tested in the evaluation. The approach aims to improve day-to-day diabetes care by allowing free-form inquiries about glucose data while maintaining strict privacy protections.