Recognizing Motion Onset During Robot-assisted Body-weight Unloading is Challenging but Seems Feasible

Abstract

Patients with neurological gait impairments fear to fall while performing locomotor tasks like sit to stand, stand to sit, and level-ground walking without assistance. This prevents them from participating in daily life. Multi-directional Body-Weight Support (BWS) systems aim to assist training in a safe environment to overcome this limitation. To ensure the safety and comfort during training, BWS systems should effectively and automatically assist patients with forces that support the desired locomotor task.Instead of manual switching of controllers through the therapist, we propose a machine learning-based motion onset recognition model that aims at automatic controller switching (4 gait related task classes in this paper). In addition to the data provided by the gait rehabilitation robot FLOAT, data of three Inertial Measurement Units (IMUs) attached to the sternum and middle of the outer thighs on 6 healthy participants were used to predict motion onset of these 4 gait-related tasks.Four train data sets have been built up from the synchronously obtained data from the IMUs and the BWS system by dividing them into observation windows with sizes of 100 ms and 200 ms and overlap factors of 50% and 90%. From each training dataset, 108 time-domain features have been extracted, ranked, and reduced using the Minimum Redundancy Maximum Relevance (MRMR) method from each train set before applying it to the classifier. The dominant features were applied for training and comparing four different classifiers. The performance of the classifiers has been evaluated based on leave one participant out cross-validation. The ensemble classifier obtained from the train data set with a window size of 100 ms and 90% overlap achieved the best performance with an F1_score of 83.7%.