This master thesis is a simulation-based study of different factors determining bias in connectivity between brain sources reconstructed using electroencephalography (EEG) techniques. In particular, the goal of this document is to propose and analyze algorithms to carry out feature ranking with different algorithms i.e. giving a measure of usefulness for each input variable.
In addition to this, we will evaluate factors determining connectivity between brain sources reconstructed from sensor measures, based on false positive rate (FPR). The idea is to have a set of features which explain false connection between brain sources (via false positive rate).
This thesis attempt to explore comparative study of state-of-the-art feature selection methods. Recursive Feature Elimination algorithm will further be compared with Univariate Feature Selection (Chi-squared statistics) and Gradient Boosting Machine. Selecting the most representative features will provide a better understanding of the underlying process.
High Performance Computer of Ghent University has been used for massive data simulation.
This Master thesis is structured as follow. Chapter 1 contains a brief overview of EEG and source localization problem. Afterwards, related work on this subject explained in detail: experiment design, performance metrics and different source localization and connectivity estimate explanation. We end up by presenting research question.
In Chapter 2 we discuss theoretical background of machine learning and statistical methods that we care going to use: Binary classification, different types of Feature selection techniques.
In Chapter 3 whole process of feature selection process is detailed from data generation, preprocessing, explorative data analysis and feature selection.
At the end, in Chapter 4 we discuss our findings and propose future works related to this thesis.