Improving Telerehabilitation in Pediatric Cerebral Palsy Using Machine Learning and Social Robots

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Cerebral palsy (CP) is the most common motor disorder in young children. There is no cure, but disciplined rehabilitation can improve outcomes. A critical component of rehabilitation is continuous assessment of patient function. Patients in rural areas can have difficulty accessing care. Telerehabilitation provides an option to extend access to care, but has limitations. The foci of this project are 1) understanding how a social robot physically co-located with the patient during a telerehabilitation assessment alters the activity by the patient, potentially leading to changes in the ability of the clinician to perform assessment and 2) whether computer vision and machine learning can be used to assess patients. Together, these two complementary goals will show a path forward for remote treatment of patients with CP and similar conditions. The effect of using a social robot in telerehabilitation will be examined through a study where pediatric CP subjects and typical subjects interact with a remote operator in three conditions: face-to-face, over traditional telepresence, and over telepresence with a social robot present. Direct changes in the level of subject interaction and compliance will be measured through surveys and video coding. The effect on quality of assessment will be measured by presenting expert therapists with first-person video recordings from each condition and comparing the variance of their grading for each condition. To truly realize the promise of using remote assessment to extend care, automated grading of assessments is necessary. To evaluate the feasibility of this, videos of children with various levels of upper extremity function along with their box and block scores and clinician ratings will be used to train two algorithms. Both algorithms will begin by using off the shelf convolutional neural network based tools to extract the pose of the subjects. The first algorithm will be hand designed. It will learn how to weight known metrics of motion, such as movement speed, time to maximum speed, and number of speed peaks, using principal component analysis and a naive Gaussian classifier. The second algorithm will use a custom neural network operating directly on the time-series pose data. Both algorithms will attempt to, given video of a novel subject, predict the level of function as would be predicted by a therapist. Both algorithms will be analyzed to discover their underlying decision-making philosophies, which may give insight into what parameters of motion clearly differentiate levels of function. The project will be done in the context of a pre-doctoral training plan. The plan focuses on developing an independent researcher at the intersection of robotics and rehabilitation science. This will be done within Mechanical Engineering, Physical Medicine and Rehabilitation, and the General Robotics, Automation, Sensing, and Perception (GRASP) laboratory at the University of Pennsylvania with additional mentorship and experience at the Children's Hospital of Philadelphia.