Tatsuji Munaka

Tokai University School of Information and Telecommunication Engineering, Tokyo,
Japan

Abstract Title:

A method of finding the early stage of Dementia using Dialogue Data

Biography:

Tatsuji Munaka is a Professor at the Graduate School of Information and Telecommunication Engineering, Tokai University. He joined Mitsubishi Electric Corporation and worked until 2015. He received a DE degree in science and engineering from Shizuoka University, Japan, in 2003. One of his current research areas is the realization of healthcare services using IoT technology.

Research Interests:

Japans rapidly aging society has made early detection and preventive intervention of dementia increasingly critical. Conventional assessments rely on questionnaires and interviews in clinical settings, creating challenges related to evaluator expertise and limited opportunities for repeated testing. Recently, approaches to detecting cognitive decline from natural conversation have gained attention; however, Japanese-language dementia dialogue datasets remain scarce, hindering the development of machine learning based evaluation models. To address this gap, we generated synthetic dementia-like dialogues using ChatGPT, focusing on five core symptoms: memory impairment, disorientation, language disorder, impaired judgment, and executive dysfunction. These dialogues were then reformatted into NLP task structures natural language inference, semantic similarity, and question answering following the JGLUE benchmark framework. A Japanese BERT model was fine-tuned on these tasks, enabling classification of 20 dementia-specific speech features, including semantic confusion, confabulation, and misrecognition of persons or time. To capture symptom progression over time, we applied time-series cross-validation and leave-one-out cross-validation (LOOCV) on one-week dialogue sequences. The time-series evaluation enabled the detection of gradual conversational changes, while LOOCV assessed generalization to unseen virtual patients.This design demonstrated the potential of the model to track cognitive decline dynamics and supported its utility for individualized monitoring. This study provides a first step toward scalable, language-specific tools for monitoring cognitive health in aging populations.