A Study on Deep Learning based Operation of Gait Rehabilitation System

Abstract

Gait rehabilitation is a critical part of the therapy plan for patients with locomotor dysfunctions in the lower limbs. The purpose of gait rehabilitation is to re-train the nervous system, re-build muscle strength, improve balance, and re-train kinematics to reduce the stresses applied to bones and muscles. Currently, gait rehabilitation is generally based on physical therapy interventions with the robotic approach still only marginally employed. Robotic devices are increasingly being accepted by many researchers and clinicians for rehabilitating physical impairments in the lower limbs. These devices provide safe, intensive, and task-oriented rehabilitation for people with mild to severe motor impairments after neurological injuries. In principle, robotic training could increase the intensity of affordable therapy, and offer advantages such as precisely controllable assistance or resistance during movements and repeatability. For robotic gait training, pre-tests are performed on all patients to identify those with conditions that complicate using robot training for rehabilitation. Setting the correct training parameters (e.g., walking speed and weight support ratio) would enhance the training efficacy and prevent potential safety hazards that could occur while training. Despite these precautionary measures, to avoid injuring patients, emergency stops are frequently required in real rehabilitation training with exoskeleton-type robotic systems. This thesis proposes a deep learning(DL)-based real-time emergency stop prediction method that, without prior anomaly knowledge, can be used in robotic gait rehabilitation training systems. The DL-based model learns and predicts the pattern of normal data and based on this result a threshold-based algorithm is developed to predict an emergency stop when abnormal behavior is observed. This thesis proposes a time-series deep model-based emergency stop prediction method that prevents the human and material loss caused by interruptions of gait rehabilitation training and emergency stops. The proposed emergency stop prediction method is based on accurately estimating the time-series trend of normaltraining by using a deep model training approach based only on normal gait rehabilitation training data. Thereafter, the difference between the estimated time series trend and the actual training data is acquired using the accumulative method. To evaluate the accuracy of the proposed methodology, the normal and abnormal training data of real patients were used. Consequently, emergency stops that occur in an actual gait rehabilitation system were shown to be detectable using the proposed emergency stop prediction method. The demand for an active gait rehabilitation program is gradually increasing among rehabilitation practitioners. However, the current gait rehabilitation systems, require additional sensors to measure bio-signals, or simply use the joint torque measured by the gait rehabilitation system without physiological feedback to measure the patient’s participation. This paper proposes a method for physiologically estimating patient participation for more active lower limb rehabilitation without the application of additional hardware. To this end, torque and EMG data were synchronized and measured under various training conditions using the gait rehabilitation system. Linear regression analysis was applied to torque-EMG pairs from the measured torque-EMG data that were found to be statistically significant through correlation analysis. Consequently, real-time estimation of muscle activity of lower limb during rehabilitation training was shown to be possible with only the torque data from the gait rehabilitation system.