Search

Hybrid brain-machine interfaces (BMIs) combining brain and muscle activity are a promising therapeutic alternative for rehabilitation of stroke patients with severe paralysis. In this study, we compare different approaches utilizing electroencephalographic (EEG) and electromyographic (EMG) activity to detect movement attempts of stroke patients with complete hand paralysis. Data of 20 patients with a chronic stroke involving the motor cortex were analyzed, and the performance of EEG-based, EMG-based or hybrid classifiers were simulated offline. We show that the combination of EEG and EMG improves the accuracy of movement detection, but that muscles unrelated to the task can also provide high accuracies, reflecting compensatory mechanisms. This result underscores the importance of appropriate designs of hybrid BMIs to maximize their rehabilitative potential.

Brain machine interfaces (BMIs) have previously been utilized to control rehabilitation robots with promising results. The design and development of more dexterous and user-friendly rehabilitation platforms is the next challenge to be tackled. We built a novel platform that uses an electro-encephalograpy-based BMI to control a multi-degree of freedom exoskeleton in a rehabilitation framework. Its applicability to a clinical scenario is validated here with six healthy subjects and a chronic stroke patient using motor imagery and movements attempts. Therefore, this study presents a potential system to carry out fully-featured motor rehabilitation therapies.

Myoelectric control constitutes a promising interface for robot-aided motor rehabilitation therapies. The development of accurate classifiers and suitable training protocols for this purpose are still challenging. In this study, eight healthy participants underwent electromyography (EMG) recordings while they performed reaching movements in four directions and five different hand movements wearing an exoskeleton on their right upper-limb. We developed an offline classifier based on a back-propagation artificial neural network (ANN) trained with the waveform length as time-domain feature extracted from EMG signals to classify discrete movements. A maximum overall classification performance of 75.54 % ± 5.17 and 67.37 % ± 8.75 were achieved for reaching and hand movements, respectively. We demonstrated that similar or better classification results could be achieved using a small number of electrodes placed over the main muscles involved in the movement instead of a large set of electrodes. This work is a first step towards a discrete decoding-based myoelectric control for a motor rehabilitation exoskeleton.

Muscle synergies can be seen as fundamental building blocks of motor control. Extracting muscle synergies from EMG data is a widely used method in motor related research. Due to the linear nature of the methods commonly used for extracting muscle synergies, those methods fail to represent agonist-antagonist muscle relationships in the extracted synergies. In this paper, we propose to use a special type of neural networks, called autoencoders, for extracting muscle synergies. Using simulated data and real EMG data, we show that autoencoders, contrary to commonly used methods, allow to capture agonist-antagonist muscle relationships, and that the autoencoder models have a significantly better fit to the data than others methods.