Phase Segmentation and Percentage Prediction of Trunk Movement Cycle

• Published in: Proceedings of the ... IEEE/RAS-EMBS International Conference on Biomedical Robotics and Biomechatronics pp. 1 - 6
• Main Authors: Ai, Xupeng, Santamaria, Victor, Prado, Antonio, Agrawal, Sunil K.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21.08.2022
• Subjects: Deep learning; Motion segmentation; Prediction algorithms; Predictive models; Supervised learning; Tracking; Training
• ISSN: 2155-1782
• DOI: 10.1109/BioRob52689.2022.9925269

Deep learning-based cyclic human motion prediction is widely used in motor learning and rehabilitation. These prediction algorithms are typically used for offline measurements. However, algorithms designed as tools to improve the efficacy of activity-based motor trainings are still being explored. This paper proposed a novel Encoder-Recurrent-Decoder structured deep neural network. Its goal is to segment the phase and predict the percentage of the cycle of sequential postural trunk movements during sitting. The model's performance was compared to three commonly used deep learning models for human motion prediction using five evaluation metrics. All models were tested in 4320 trials in 45 healthy subjects. The novel Encoder-Recurrent-Decoder model outperformed the competing models. Additionally, it can predict trunk movement cycle percentage within a Mean Absolute Error (MAE) of 7% and reach 89% accuracy in trunk phase segmentation. In future studies, the trunk movement prediction outcomes will be used in the design of a high-level controller for a robotic platform to maximize machine-clinician interaction during the delivery of motor interventions.