A deep learning model for assistive decision-making during robot-aided rehabilitation therapies based on therapists' demonstrations

BackgroundA promising approach to improving motor recovery during rehabilitation is the use of robotic rehabilitation devices. These robotic devices provide tools to monitor the patient's recovery progress while providing highly standardized and intensive therapy. A major challenge in using these robotic devices is the ability to decide when to assist the user. In this context, we propose a Deep Learning-based solution that can learn from a therapist's criteria when a patient needs assistance during robot-aided rehabilitation therapy.MethodsAn experimental session was conducted with diverse people who suffered from neurological conditions. The participants used an upper limb rehabilitation robot to play a point-to-point game. A therapist supervised the robot-aided rehabilitation exercises and assisted the participants when considered necessary. This assistance provided by the therapist was detected to label those trajectories that were assisted to train a Deep Learning model that learns from the therapist when to assist. A series of transformations have been applied to the trajectories performed by the participants to generalize the method. Furthermore, the trajectory data was divided into sequences to be introduced to the model and continuously infer whether the user needs assistance. The data acquired during the experimental sessions have been divided into two datasets to train and evaluate the model: intra-participants (80% training, 20% validation) and test participants. The architecture of the Deep Learning model is conceived to perform time-series classification. It consists of diverse one-dimensional convolutional layers, a convolutional attention mechanism, and a Global Average Pooling layer. In addition, the output layer has one neuron with the sigmoid activation function, whose output can be interpreted as a probability of assistance. The model proposed in this study has been evaluated according to different metrics. In addition, the impact of applying fine-tuning to adapt the assistance to each user has been evaluated with the test participants.ResultsThe proposed model achieved an accuracy of 91.39% and an F1-Score of 75.15% with the validation dataset during a sequence-to-sequence evaluation, surpassing other state-of-the-art architectures. When evaluating the trajectories collected in the test dataset, the method proposed achieved an accuracy of 76.09% and an F1-Score of 74.42% after applying fine-tuning to each participant.ConclusionsThe results achieved by our Deep Learning-based method show the feasibility of learning assistance decision-making from experimented therapists. Furthermore, fine-tuning can be applied to personalize the assistance to each user and improve the accuracy of the method presented when deciding whether to assist with the rehabilitation robot.