Recent demographic trends have caused a shift in long-term care from institutions to home- and community-based services. This shift favors telehealth as an alternative to conventional therapy, however with various limitations in their framework design. To address these limitations, we propose the use of latent variable models for learning therapist-induced motion.
In this presentation, we propose modifications to two latent variable models and extract meaningful features from the modified latent spaces. First, we introduce an aggregation approach to muscle synergy extraction. Second, we introduce a reorganization approach to a Gaussian Process embedding of multiple motor task performances. The utility of the proposed metrics is demonstrated through the sit-to-stand task, an important whole-body exercise in physiotherapy. By comparing the extracted metrics through various natural and expert-induced conditions, we are able to gain some insight on the expert strategy as well as build tools for providing expert-level feedback.