Generating Synthetic Behavioral Health Data: A Multi-Agent LLM System

Mobile health (mHealth) tools offer a scalable solution for delivering behavior change interventions. In recent years, personalization has been increasingly adopted in mHealth interventions. However, developing personalized models requires user data, which is often scarce or unavailable for new interventions. This problem is often addressed by creating health behavior personas, which are virtual representations of users. Personas enable the creation of simulation environments that allow researchers to configure and train intervention models before deployment. However, developing simulation models to generate synthetic data for health behavior personas requires interdisciplinary expertise that is often inaccessible or time-intensive for domain researchers. To address this challenge, we introduce a novel cooperative multi-agent architecture based on large language models (LLMs) to empower domain experts to create and simulate health behavior personas through natural language interaction. A proof-of-concept is offered through a case study on literature-based smoking cessation personas, which shows how our approach reduces barriers for researchers while ensuring reliable synthetic behavioral health data.