Application of TimeGAN to IMU-based Data of Upper Limb Range of Motion

Time-series generative adversarial networks (TimeGAN) were recently developed to produce synthetic time-series data for varied applications. Most prior uses of TimeGAN in biomechanics and rehabilitation research did not consider data from inertial measurement unit (IMU) sensors for upper limb range of motion (ROM), especially in the context of disability simulation studies. In this paper, we used TimeGAN to produce synthetic three-dimensional (3D) ROM data for elbow flexion-extension movement performed under four experimental conditions within a disability simulation framework. In each experimental condition, we collected 3D ROM data of human subjects using the wearable inertial sensors for exergame (WISE) system and then produced synthetic ROM data using TimeGAN. Our goal was to produce accurate synthetic data to circumvent difficulties related to capturing data from a large number of human subjects. To assess the performance of TimeGAN in producing synthetic data, we used principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and discriminative score. The results show that TimeGAN produced synthetic ROM data of upper limb movements quite well. Specifically, the average discriminative score in the unrestricted, restricted shoulder, restricted elbow, and restricted forearm conditions was found to be 0.12, 0.08, 0.18, and 0.11, respectively.Clinical Relevance-This work illustrates the use of TimeGAN to produce synthetic 3D ROM data. A large 3D ROM dataset can help in training deep learning models to classify impairments in various upper limb joints, guide therapy interventions, and assess rehabilitation progress for patients suffering from movement ailments such as hemiplegia.