Integration of large language models and sensor data in recommendation systems for personalized physiological and dietary monitoring

Abstract

This study proposes the «Formal -of-Thought» (FoT)—an architecture that shifts artificial intelligence from passive generation to active formal reasoning. This approach uses the Health-LLM principle to convert numerical sensor data in the form of time series into textual descriptions based on an individual user profile. It combines this with the PHIA agent, which utilizes multi-step ReAct reasoning, Python code generation (Pandas), few-shot learning via embedded sentence-T5 and K-means clustering, as well as the GPT-4 Text -to-SQL method with few-shot learning and RAG for heterogeneous data.