Multi-agentic Framework for Developing and Evaluating Biometric-based Multimodal Synthetic Personas: A Case Study on Pregnancy and Postpartum Users

Despite the rise of LLM-based health agents, robust and privacy-preserving evaluation frameworks remain scarce. We present a framework using a Transformer-based TimeGAN to generate synthetic biometric time-series from 3,497 real participant data. These time series are integrated with LLM-generated narrative vignettes to create biometrically-aligned multimodal personas for realistic user-agent simulations. Employing an LLM-as-a-Judge methodology, we evaluate agent performance across Passive and Proactive modes. Our results demonstrate that this framework effectively identifies critical gaps in current LLM capabilities regarding factual correctness, data usage, and clinical plausibility. While focused on the postpartum period, our methodology provides a scalable tool for optimizing and evaluating personal health agents across diverse biometric and clinical applications.