Detection of dementia from responses to atypical questions asked by embodied conversational agents

Detection of dementia requires examinations, such as blood tests and functional magnetic resonance imaging (fMRI), that can be very stressful for the patient. Previous studies proposed screenings for easy detection of dementia that utilized acoustic and language information derived from conversations between patients and medical staff. Although these studies demonstrated effectiveness in automatically detecting dementia, the tasks used were created based on neuropsychological tests. The effect of habituation on this limited variety of tasks might have a negative impact on routine dementia screening. We propose a method to detect dementia using responses to more atypical questions asked by embodied conversational agents. Through consultations with neuropsychologists, we created a total of 13 questions. The embodied conversational agent obtained answers to these questions from 24 participants (12 dementia and 12 non-dementia). We recorded their responses and extracted speech and language features. We classified the two groups (dementia/non-dementia) by a machine learning algorithm (support vector machines and logistic regression) using the extracted features. The results showed a 0.95 detection performance in the area under the curve of the receiver operating characteristic (AUROC). This result demonstrates that our system using atypical questions can detect dementia.