Deep learning for gait prediction: an application to exoskeletons for children with neurological disorders

Cerebral Palsy, a non-progressive neurological disorder, is a lifelong condition. While it has no cure, clinical intervention aims to minimise the impact of the disability on individuals' lives. Wearable robotic devices, like exoskeletons, have been rapidly advancing and proving to be effective in rehabilitating individuals with gait pathologies. The utilization of artificial intelligence (AI) algorithms in controlling exoskeletons, particularly at the supervisory level, has emerged as a valuable approach. These algorithms rely on input from onboard sensors to predict gait phase, user intention, or joint kinematics. Using AI to improve the control of robotic devices not only enhances human-robot interaction but also has the potential to improve user comfort and functional outcomes of rehabilitation, and reduce accidents and injuries. In this research study, a comprehensive systematic literature review is conducted, exploring the various applications of AI in lower-limb robotic control. This review focuses on methodological parameters such as sensor usage, training demographics, sample size, and types of models while identifying gaps in the existing literature. Building on the findings of the review, subsequent research leveraged the power of deep learning to predict gait trajectories for the application of rehabilitative exoskeleton control. This study addresses a gap in the existing literature by focusing on predicting pathological gait trajectories, which exhibit higher inter- and intra-subject variability compared to the gait of healthy individuals. The research focused on the gait of children with neurological disorders, particularly Cerebral Palsy, as they stand to benefit greatly from rehabilitative exoskeletons. State-of-the-art deep learning algorithms, including transformers, fully connected neural networks, convolutional neural networks, and long short-term memory networks, were implemented for gait trajectory prediction. This research presents findings on the performance of these models for short-term and long-term recursive predictions, the impact of varying input and output window sizes on prediction errors, the effect of adding variable levels of Gaussian noise, and the robustness of the models in predicting gait at speeds within and outside the speed range of the training set. Moreover, the research outlines a methodology for optimising the stability of long-term forecasts and provides a comparative analysis of gait trajectory forecasting for typically developing children and children with Cerebral Palsy. A novel approach to generating adaptive trajectories for children with Cerebral Palsy, which can serve as reference trajectories for position-controlled exoskeletons, is also presented.