Development of an EEG-based recurrent neural network for online gait decoding

Recent neuroscientific literature has shown that the use of brain-controlled robotic exoskeletons in walking rehabilitation induces neuroplasticity modi- fications, possibly leading to a higher likelihood of recovery and maintenance of lost motor functions due to a neural lesion, with respect to traditional re- habilitation. However, the gait decoding from brain signals remains an open challenge. The aim of this work is to implement and validate a deep learning model for online gait decoding that exploits Electroencephalography (EEG) infor- mation to predict the intention of initiating a step, which could be used to trigger the assistance of a lower-limb exoskeleton. In particular, the model exploits a Gated Recurrent Units (GRU) deep neural network to handle the time-dependent features which were identified by analysing the neural cor- relates preceding the step onset (i.e., Movement-Related Cortical Potentials (MRCP)). The network was evaluated on a pre-recorded dataset of 11 healthy subjects walking on a treadmill. The network’s architecture (e.g., number of GRU units) was optimized through grid search. In addition, to deal with the data scarcity problem of neurophysiological applications, I proposed a data augmentation procedure to increase the dataset available to train the model of each subject. With the proposed approach, the model achieved an average accuracy in detecting the step onset of 89.7 ± 7.7% with just the 15% of the dataset for each subject (∼70 steps), and up to 97.8 ± 1.3% with the whole dataset (∼440 steps). This thesis support the use of a memory-based deep learning model to de- code walking activity from non-invasive brain recordings. In future works, this model will be exploited in real time as a more effective input for devices restoring locomotion in impaired people, such as robotic exoskeletons.