Motor imagery based brain-computer interface: improving the EEG classification using Delta rhythm and LightGBM algorithm

This article contains a new method to improving the EEG motor imagery classification system quality with an application on BCI competition IV 2a, 2b, and PhysioNet EEG-MI datasets. This work uses a bandpass filter to eliminates all unused signals and then increases the prediction accuracy from 50% to more than 96% in both binary and multi-class cases, knowing that applying PSO optimizer on the parameters of the LightGBM classifier allows to find the best and stable status of EEG signals classification, also decision tree algorithm (DT) allows to get the importance degree of all acquisition electrodes used in the classification stage. This work also uses the correlation matrix to determined all artifacts between different electrodes, in such a way the prediction accuracy value increases from 50% and 60% to higher values of 96% and 98% in binary and multi-class classification, and high prediction speed remains more than 63703 and 2395 samples per second in binary and multi-class cases respectively. A comparison at the end of related works found a maximum accuracy value of around 85.5%.