A Multi-Agent Dual Dialogue System to Support Mental Health Care Providers

TL;DR

This paper addresses clinician workload and access gaps in mental health care by proposing a human-in-the-loop dual-dialogue system that augments therapists without replacing them. It employs a modular multi-agent LLM architecture with retrieval-augmented resources to propose responses, extract themes, summarize dialogues, and guide content tailored to the client. Empirical evaluation shows LLM-generated responses achieve empathy ratings comparable to professional therapists, suggesting practical integration into clinical workflows while highlighting remaining concerns around privacy, bias, and multilingual performance. The work demonstrates potential for scalable, therapist-guided support tools but calls for further validation, privacy-preserving deployments, and broader community adoption.

Abstract

We introduce a general-purpose, human-in-the-loop dual dialogue system to support mental health care professionals. The system, co-designed with care providers, is conceptualized to assist them in interacting with care seekers rather than functioning as a fully automated dialogue system solution. The AI assistant within the system reduces the cognitive load of mental health care providers by proposing responses, analyzing conversations to extract pertinent themes, summarizing dialogues, and recommending localized relevant content and internet-based cognitive behavioral therapy exercises. These functionalities are achieved through a multi-agent system design, where each specialized, supportive agent is characterized by a large language model. In evaluating the multi-agent system, we focused specifically on the proposal of responses to emotionally distressed care seekers. We found that the proposed responses matched a reasonable human quality in demonstrating empathy, showing its appropriateness for augmenting the work of mental health care providers.

Figures (6)

• Figure 1: Schematic of dual dialogue system design.
• Figure 2: Schematic of system architecture of dual dialogue system.
• Figure 3: The therapist's front-end user interface showcases the two concurrent conversations and AI features, including "Propose response," "Recommend resources," and open-ended chat.
• Figure 4: Examples of functionalities, including proposing responses, extracting relevant resources, and analyzing and summarizing conversations.
• Figure 5: Histograms of rating items for responses generated by GPT-4o (first row), Llama 3 70b (second), Llama 3 8b (third), and professional therapists (last), as evaluated by GPT-4o.