P1‐559: TOWARD PERCEPTUAL DEEP LEARNING–BASED ABNORMAL BEHAVIOR PATTERN ANALYSIS FOR PATIENTS WITH ALZHEIMER'S DISEASE
Jeonghwan Gwak, Jong‐In Song, Kiseon Kim, Moongu Jeon, Cheolbin Park, Hyunsu Jeong, Kyu Yeong Choi, Jung Eun Park, Jung Sup Lee, Byeong Chae Kim, Kun Ho Lee · Alzheimer s & Dementia · 2018
As severity of AD progresses, behavior and cognition capabilities are both decreased and they are highly correlated. It is recognized that behavior patterns of Alzheimer's disease (AD) can be different (sometimes categorized) for different stages although they have personal characteristics. This work focuses on devising a deep learning-based analysis method of different abnormal behavior patterns using multiple visual sensors. The experiments consist of 4 types of balance tests and 4 types of gait tests with complex dual tasks (which is to measure walking ability while performing attention/cognition tasks during the tests). The proposed method is based on a temporal deep learning approach and has three key steps: 1) modelling motion patterns in a short time (e.g., 1.5 seconds or 45 frames per seconds) using body-part segmented regional activity analysis, 2) extracting 3D motion features, and 3) analyzing motion patterns using our proposed temporal deep learning framework with a perceptual residual error criterion. In the experiments, we used 25 cognitively normal (CN) subjects and 10 prodromal AD (pAD) subjects. As the evaluation measure, voxel-based entropy was used to investigate the abnormal behavior patterns. From the resulting analysis, we could observe that the abnormality deviation of pAD subjects from CN subjects has 17.19% (with standard deviation of 2.74%) on average. The results indicate that the abnormal patterns can be effectively used to classify the two subject groups.