Abstract
This thesis proposes a method for mathematical modeling of human movements by using deep artificial neural networks, with application in modeling patient exercise episodes performed during physical therapy and rehabilitation sessions. The generative adversarial network (GAN) structure is adopted, whereby a discriminative and a generative model are trained concurrently in an adversarial manner. The capacity of GAN models for generating synthetic data offers a potential to artificially augment the size of datasets for biomedical applications, where collecting large datasets is notoriously challenging, due to the need for access to patients, as well as due to privacy, safety, and ethics concerns. Synthetically augmented datasets have demonstrated improved robustness and overall performance of machine learning models across various data formats and modalities. The thesis examines different network architectures, with the discriminative and generative models structured as deep subnetworks of hidden layers comprised of convolutional or recurrent computational units. The models are validated on a dataset of physical rehabilitation movements recorded with an optical motion tracker. The results demonstrate an ability of GAN network architectures for generation of movement examples that resemble the recorded rehabilitation movement sequences, and for classification of unseen instances of the movements.