Classification methods for eeg patterns of imaginary movements

The review focuses on the most promising methods for classifying EEG signals for non-invasive BCIs and theoretical approaches for the successful classification of EEG patterns. The paper provides an overview of articles using Riemannian geometry, deep learning methods and various options for preprocessing and "clustering" EEG signals, for example, common-spatial pattern (CSP). Among other approaches, pre-processing of EEG signals using CSP is often used, both offline and online. The combination of CSP, linear discriminant analysis, support vector machine and neural network (BPNN) made it possible to achieve 91% accuracy for binary classification with exoskeleton control as a feedback. There is very little work on the use of Riemannian geometry online and the best accuracy achieved so far for a binary classification problem is 69.3% in the work. At the same time, in offline testing, the average percentage of correct classification in the considered articles for approaches with CSP – 77.5$\pm$5.8%, deep learning networks – 81.7$\pm$4.7%, Riemannian geometry – 90.2$\pm$6.6%. Due to nonlinear transformations, Riemannian geometry-based approaches and complex deep neural networks provide higher accuracy and better extract of useful information from raw EEG recordings rather than linear CSP transformation. However, in real-time setup, not only accuracy is important, but also a minimum time delay. Therefore, approaches using the CSP transformation and Riemannian geometry with a time delay of less than 500 ms may be in the future advantage.