On the Use of Localized Strategies in Inverse Source Powered Brain-Computer Interfaces

A Brain-Computer Interface (BCI) is a unidirectional communication system used to input predefined control commands to an external apparatus without using conventional motor output pathways but only monitoring the brain activity. A typical BCI system is composed of three main subsystems: (i) a brain acquisition system, usually Electroencephalography (EEG) or Magnetoencephalography (MEG); (ii) a signal processing system that includes pre-processing of the EEG/MEG signals, feature extraction and the translation of the features in commands by using a machine learning (ML) classification stage; (iii) a computer controlled by the generated commands providing a visual feedback. In this thesis we focus on a specific family of BCI that leverage on EEG Source Imaging (ESI) to map the EEG recording to intracranial currents by using electromagnetic models of the brain. ESI leverages on two different electromagnetic problems to reconstruct the brain activity: the forward problem (FP) and the inverse problem (IP). The FP maps the current distribution inside the brain to the electric potential on the scalp and, in particular, at the electrodes locations. The IP performs the dual operation through multiple evaluations of the FP: given the potentials recordings on the scalp, it infers the volumetric current distribution that better fits the electrode measurements. One of the main challenges of this approach is that IP is an ill-posed problem that does not admit a unique and stable solution. Several BCI paradigms have been presented in the literature depending on the type of command to be classified and the mental strategy used to generate them. In this thesis a particular attention has been given to motor-imagery (MI) classification since it is currently the most challenging but most promising paradigm and is currently the subject of numerous academic investigations. This thesis focuses on the signal processing sub-system and, in particular, on developing a different strategy for processing and classification of the EEG-data. The specific objective of this work is to design, implement and test a new ESI-based BCI pipeline that leverages on a localized inverse method to solve the IP. This family of inverse techniques has scarcely been studied in the literature but deserves more attention. The main advantages of using localized inverse solution is the chance of obtaining a small number of freely moving equivalent dipoles summarizing the whole cerebral activity for a given measurement sample. Using such a parametric method is expected to increase the precision of the reconstructed brain activity and therefore improve the system's global classification accuracy. In order to generate an EEG dataset and test the newly developed pipelines, the simBCI framework is employed. Although the usage of simulated EEG-data helps to avoid some of the problems coming from real measurements, a careful selection of many hyper-parameters is necessary to correctly test any pipeline. For this reason, part of the thesis work is also focused on the correct tuning of the pipeline's basic values. Comparison with some of the state of the art BCI systems have been performed in order to verify if the usage of localized inverse methods deserves more research efforts. The numerical results obtained through these simulations will be presented in the last part of the document.