A study on dementia detection method with stroke data using anomaly detection

Increasing the number of elderly persons who have dementia, this is one of the severe social problems in Japan. According to the report published by the Ministry of Health, Labor and Welfare, the number of elderly persons with dementia will be around five million in 2015. This report indicates that early detection and prevention of dementia is essential. From viewpoints of early detection of dementia, the most problem is the limitation of test contents and the difficulty of taking a dementia check test on a daily basis. To solve these problems, the authors focus on drawing test using a tablet terminal to develop a dementia detection system, which can be adapted to various drawing contents including digits, characters, and pictures for increasing of dementia screening opportunity. It is, however, difficult to collect sufficient data to build the system because there are many subtypes of dementia. From this background, this position paper discusses an unsupervised anomaly detection method using healthy data only, and also aim to propose a system that gives the probability of being dementia (or other sicknesses) based on the differences from the data of healthy cases. As the first step of this study, we discuss the possibility of a dementia detection method using Variational Autoencoder.