1534-4320 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2019.2953707, IEEE Transactions on Neural Systems and Rehabilitation Engineering IEEE Transactions on Neural Systems and Rehabilitation Engineering (Manuscript) 1 Abstract—Analysis of joint motion data (AJMD) by Kinect, such as velocity, has been widely used in many research fields, many of which focused on how one joint moves with another, namely bivariate AJMD. However, these studies might not accurately reflect the motor symptoms in patients. The human body can be divided into six widely accepted parts (head, trunk and four limbs), which are interrelated and interact with each other. Therefore, in this study we attempted to investigate how the major joints of one body part move with the ones in another body part, namely multivariate AJMD. For method illustration, the motion data of sit-to-stand-to-sit for healthy participants and people with Parkinson disease (PD) were employed. Four types of multivariate AJMD were investigated by eigenspace-maximal-information- canonical-correlation-analysis, which obtained maximal- information-eigen-coefficients (MIECes), the parameters for quantifying the correlation between two sets of joints located in two different body parts. The results show that healthy participants have significantly higher MIECes than the PD patients (p-value < 0.0001). Furthermore, the area under the receiver operating characteristic curve value for the classification between healthy participants and PD patients reaches up to 1.00. In conclusion, we demonstrated the possibility of using multivariate AJMD for motion feature extraction, which may be helpful for medical research and engineering. Index Terms—bivariate analysis, human motion analysis, joint motion data, Kinect, motion feature extraction, multivariate analysis, Parkinson’s disease (PD), sit-to-stand-to-sit, velocity Peng Ren and Jorge Francisco Bosch Bayard contribute equally in this paper. Peng Ren, Li Dong, Jinying Chen, Lu Mao, Maria L. Bringas, Dezhong Yao, Marjan Jahanshahi and Pedro A. Valdes-Sosa are with the Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, University of Electronic Science and Technology of China, Chengdu, China, 610054. Email: {pren28@uestc.edu.cn, pedro@uestc.edu.cn} Jorge Francisco Bosch Bayard is with the Institute for Neurobiology, National Autonomous University of Mexico, Queretaro, Mexico, 76230. Dan Ma is with the Department of Rehabilitation Medicine, Key Laboratory of Birth Defects and Related Diseases of Women and Children, West China Second University Hospital of Sichuan University (WCSUH-SCU), Ministry of Education, Chengdu, China, 610041. Mario Alvarez Sanchez is with Dr. Georges-L.-Dumont University Hospital Centre (DGLDUHC), 330 University Avenue, Moncton, Canada, E1C2Z3. Daniel Mondejar Molejon is with Cuban Neuroscience Center, Ave 25 #15202 esquina 158, Cubanacan, Playa, Cuba. Marjan Jahanshahi is with Department of Clinical and Movement Neuroscience, UCL Queen Square Institute of Neurology, London, UK, WC1N3BG. I. INTRODUCTION A. Description of Human Motion Analysis As part of artificial intelligence research, human motion analysis has been a prominent domain of research in the last three decades, which has been widely used in many areas such as healthcare, smart home, surveillance systems, human-computer interaction, gaming and so on [1- 6]. Usually there are three research categories in human motion analysis: human gesture/action/activity recognition, human motion tracking, and the movement analysis of body and body parts. First, according to the definition in the literature [7] and [8], human motion is conceptually categorized into three levels including gesture, action and activity [9-13]. Gestures are the atomic components that describe the meaning of the motions such as “raising an arm”. Actions are single person behavior that may be consisting of multiple gestures planned temporally such as “walking”. Activities are complex sequences of actions performed by several persons. Human gesture/action/activity recognition has many applications. For example, the accelerometer and gyroscope embedded in the smartphone are utilized to detect and characterize the physical actions of our daily life for the purpose of evaluation of living habits [14]. Games are developed for deaf children that are based on recognizing American sign language [15]. Second, the basic idea behind human motion tracking is that video frames can be analyzed in order to follow the position and action of an individual through time, which can be carried out in two or three dimensions. Depending on the complexity of analysis, representations of the human body range from basic stick figures to volumetric models. Tracking generally depends on the correlation of image features between successive frames of video and taking the information such as position, color, shape, and texture into consideration [16], [17]. Third, motion analysis of the body and body parts in the medical field is very critical, which is also the focus of our article. For example, many clinical studies investigated the parameters extracted from the motion data of patients in order to accurately quantize their symptoms [18], [19]. Thus, it can be seen that human motion analysis plays an increasingly important role in many fields. Due to the active research areas and rapid development of artificial intelligence, compared with traditional movement measure methods, human motion analysis through RGB (Red, Green and Blue) camera, depth camera and wearable devices has gained increasing popularity in recent years [20-24]. RGB is a color model based on the additive synthesis, which allows three primary colors of mixed light to represent color. Most Peng Ren, Jorge F. Bosch Bayard, Li Dong, Jinying Chen, Lu Mao, Dan Ma, Mario Alvarez Sanchez, Daniel Mondejar Molejon, Maria L. Bringas, Dezhong Yao, Marjan Jahanshahi, Pedro A. Valdes-Sosa Multivariate Analysis of Joint Motion Data by Kinect: Application to Parkinson’ s Disease