LLM Digital Twin Framework Detects MCI via Conversational Patterns

A language-based digital twin uses LLMs with stylometric cues to mimic elderly individuals' conversational behavior, evaluated via a multi-head cVAE that jointly measures reconstruction quality and predicts MoCA cognitive scores. Tested on the I-CONECT dataset, the framework matches real-data reconstruction error while outperforming baselines on cognitive score prediction.