DemMA: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation

TL;DR

DemMA addresses the challenge of simulating dementia conversations by grounding personas in clinically informed subtypes and memory states, and by integrating nonverbal cues into a multimodal dialogue model. It introduces a five-agent workflow plus a Chain-of-Thought distillation strategy to produce long-horizon, low-latency, persona-consistent interactions without real patient data. The authors release DemMA-Dialogue, a synthetic dementia dialogue dataset validated by experts, and demonstrate that DemMA outperforms baselines on simulation fidelity, dialogue quality, and educational usefulness. The work advances safe, scalable training of dementia-care simulations with potential applications in caregiver training and clinical training.

Abstract

Simulating dementia patients with large language models (LLMs) is challenging due to the need to jointly model cognitive impairment, emotional dynamics, and nonverbal behaviors over long conversations. We present DemMA, an expert-guided dementia dialogue agent for high-fidelity multi-turn patient simulation. DemMA constructs clinically grounded dementia personas by integrating pathology information, personality traits, and subtype-specific memory-status personas informed by clinical experts. To move beyond text-only simulation, DemMA explicitly models nonverbal behaviors, including motion, facial expressions, and vocal cues. We further introduce a Chain-of-Thought distillation framework that trains a single LLM to jointly generate reasoning traces, patient utterances, and aligned behavioral actions within one forward pass, enabling efficient deployment without multi-agent inference. Extensive evaluations with experts, medical students, and LLM judges demonstrate that DemMA significantly outperforms strong baselines across multiple metrics.

Figures (5)

• Figure 1: DemMA integrates three components: a) Clinically grounded patient persona formation module; b) Multi-agent dialogue dataset generation pipeline encompassing memory analysis, planning, language generation, and action simulation; and c) CoT distillation multi-task training workflow for DemMA agent.
• Figure 2: LLM judgment (GPT-5.2-pro) Performance across four dementia subtypes.
• Figure 3: Scaling Effect of training dataset.
• Figure 4: Confusion matrices for dementia subtype identification (experts vs. students) across nine subtypes.
• Figure 5: Pairwise win/tie/loss on Authentic. Win rate is computed as $\mathrm{Win}/(\mathrm{Win}+\mathrm{Loss})$, excluding ties.