EEG and ECoG features for Brain Computer Interface in Stroke Rehabilitation

Abstract

The ability of non-invasive Brain-Computer Interface (BCI) to control an exoskeleton was used for motor rehabilitation in stroke patients or as an assistive device for the paralyzed. However, there is still a need to create a more reliable BCI that could be used to control several degrees of Freedom (DoFs) that could improve rehabilitation results. Decoding different movements from the same limb, high accuracy and reliability are some of the main difficulties when using conventional EEG-based BCIs and the challenges we tackled in this thesis. In this PhD thesis, we investigated that the classification of several functional hand reaching movements from the same limb using EEG is possible with acceptable accuracy. Moreover, we investigated how the recalibration could affect the classification results. For this reason, we tested the recalibration in each multi-class decoding for within session, recalibrated between-sessions, and between sessions. It was shown the great influence of recalibrating the generated classifier with data from the current session to improve stability and reliability of the decoding. Moreover, we used a multiclass extension of the Filter Bank Common Spatial Patterns (FBCSP) to improve the decoding accuracy based on features and compared it to our previous study using CSP. Sensorimotor-rhythm-based BCI systems have been used within the same frequency ranges as a way to influence brain plasticity or controlling external devices. However, neural oscillations have shown to synchronize activity according to motor and cognitive functions. For this reason, the existence of cross-frequency interactions produces oscillations with different frequencies in neural networks. In this PhD, we investigated for the first time the existence of cross-frequency coupling during rest and movement using ECoG in chronic stroke patients. We found that there is an exaggerated phase-amplitude coupling between the phase of alpha frequency and the amplitude of gamma frequency, which can be used as feature or target for neurofeedback interventions using BCIs. This coupling has been also reported in another neurological disorder affecting motor function (Parkinson and dystonia) but, to date, it has not been investigated in stroke patients. This finding might change the future design of assistive or therapeuthic BCI systems for motor restoration in stroke patients.