Design and Implementation of an LLM-Powered Personalized Health Recommender System with Wearable Data Integration

Encouraging a healthy lifestyle and avoiding chronic diseases is a persistent public health goal. Particularly, when societies with considerably higher living standards are also affected by sedentary norms. Taking forward the existing Health App project, this thesis enriches the functionality of the app with a Large Language Model-powered personalized recommendation generator. In this thesis, the challenge of developing an engine that provides applicable, comprehensive, and individualized health and wellness advice is addressed. A cloud ready and modular recommender system is designed and deployed which can assimilate seamlessly with the already existing Health App’s mobile platform. By exploiting heterogeneous user data which includes metrics and logs related to physical activity, sleep, nutrition, and stress which is either acquired by Fitbit, and manually entered by the user, the recommender system yields customized guidance to enhance the health and well-being of the user. Meanwhile, the feasibilities of real world data ingestion and flexible API deployment was sufficiently demonstrated. The fine-tuning of Meta’s LlaMA2 model using Low-Rank Adaptation (LoRA) method remains central to this thesis. The fine-tuning was performed on a curated dataset that simulated true-to-life doctor-patient conversations. The dataset was balanced and curated based on the information and knowledge from the book: “Outlive: The Science and Art of Longevity – Peter Attia, MD.” Data engineering, crafting conversational prompts, training the model on GPU clusters, and implementing a resilient cloud-based API in combination with real-time inference hosted on RunPod – a cloud infrastructure, is what encompasses the technical pipeline of this project. To evaluate the model, automated metrics including BLEURT and BERTscore were employed in combination with G-Eval which is a qualitative review of the generated recommendations, to assess the helpfulness, accuracy, empathy and relevance to the user query. As demonstrated by the experimental findings, the recommender system has the ability to provide personalized, safe, and most importantly medically relevant well-being advice, thus manifesting its potential to serve as a virtual assistance tool for individual users, whether it be patients or healthcare professionals. Crucial challenges included managing diverse data and adaptation of model. Along with documenting the design and implementation of an LLM-powered wellness recommender, this thesis also yields realistic discernment into integrating AI-based customization in mobile health applications and can act as a foundation for current research and clinical verification.