Deterministic and Machine Learning Strategies in High Resolution Electroencephalography for BCI Applications

Brain-computer interfaces (BCI) are systems that translate the brain activity of their users into commands for a machine, thus allowing communication with the external world bypassing the motor-nervous system. BCIs are used in a wide range of scenarios including the assistance of people with serious neuromuscular disorders, such as the locked-in syndrome and the amyotrophic lateral sclerosis, and the development of new kinds of human-machine interfaces. The electroencephalography (EEG) is one of the most popular brain activity acquisition modality in this field, thanks to its low cost and high portability. However, the processing of the EEG signal involves some difficulties: the signal has a quasi-stochastic nature and the same mental tasks have different manifestations between subjects because of dissimilarities in physiology, anatomy and mental strategy. In addition, the reconstruction of volumetric current distribution inside the brain that generates the EEG signals is challenging and, in particular, a high number of electrodes is needed (HR-EEG) for this so called EEG Source Imaging (ESI). Thus, complex signal processing and elaborate classification strategies are necessary. Over the years, many algorithms have been proposed, most of them involve tuning of numerous hyper-parameters. In practice obtaining statistically significant experiments using these algorithms requires a large population of subjects to perform long and fatiguing BCI sessions. To overcome these limitations, simulated data have recently started attracting the attention of the community because they can be used for relatively fast and inexpensive testing of research hypotheses and pipelines with minimum effort, before moving to real data. The objective of this thesis is to determine whether or not ESI based BCIs can reach or even outperform purely machine learning based state-of-the-art techniques. While it is expected that both approaches can outperform each other in specific scenarios, this work investigates in depth the conditions under which one or the other performs better. Because of the very large number of parameters to study – either algorithm-specific or physiological – an open-source framework called simBCI has been used for generating artificial EEG data and to test various classification algorithms. Two BCI pipelines based on Riemannian Geometry have been implemented and included in simBCI. The first one, called Minimum Distance to Riemannian Mean, uses the concepts of Riemannian Distance and Riemannian Mean to classify EEG trials described as Spatial Covariance Matrices (SCM). The second one, called Tangent Space LDA, projects the SCM matrices from the Riemannian manifold to one of its infinite tangent spaces, where linear techniques can be applied. These pipelines are compared with the historical BCI state of the art, the CSP-LDA algorithm, and with more recent ESI-based pipelines. In these experiments, the influence of different experimental parameters and numerical brain models has been evaluated. Finally, our hypothesis are validated on real data obtained from BCI competition IV. The overall experimental protocol and the results obtained are presented throughout this thesis.