An Interactive Monitoring Robot for Dementia Mitigation via Daily Conversations with Multiple LLMs

Abstract

This study presents the development of an interactive cognitive function assessment system designed for everyday use, to address the growing number of dementia cases and the rising burden of care in an aging society with a declining birth rate. The proposed system integrates multiple large language models (LLMs) to achieve this goal. The system consists of three distinct LLM modules: (1) Detector responsible for identifying user intentionality, (2) ChitChat, facilitating casual conversation, and (3) StrictTask, managing task-oriented dialogues. These modules are orchestrated using the LangChain architecture to function as a cohesive assessment system. Evaluation experiments were conducted to assess the accuracy of the HDS-R (Hasegawa's Dementia Scale-Revised) in the proposed LLM-based system and compare it with existing rule-based systems. The evaluation employed pre-designed conversational scenarios as test cases, identifying specific errors in each scenario. Results indicated that the proposed system demonstrated flexible handling of user interruptions and adaptable responses to user inputs, while also probabilistically managing the strict conditional branches dictated by HDS-R. In comparative evaluations, the proposed system exhibited lower accuracy in strict conditional branching compared to the existing rule-based systems. However, it significantly outperformed the conventional systems in terms of dialogue quality, offering features such as flexible responses, immediate reactions, and engaging chit-chat. These features improve user engagement and overall system usability, demonstrating the advantages of the proposed system as a user-friendly cognitive assessment tool.