Motor Imagery Tasks Based Electroencephalogram Signals Classification Using Data-Driven Features

Abstract

Brain-Computer Interface (BCI) system consist of a variety of different applications based on the processing of electroencephalograph (EEG). One of the most significant categories are based on EEG signals segmentation for “Motor Imagery” (MI) classification.

When analytic methods use a fixed set of basis functions, the EEG signals frequently exhibit poor time-frequency localization. Additionally, these signals have a low signal-to-noise ratio (SNR) and highly non-stationary characteristics. As a result, BCI systems frequently have high error rates and low task detection accuracy.

This work is aiming to introduce the adaptive and data-driven based feature extraction method for MI-tasks classification. In this regard, empirical mode decomposition (EMD) and ensemble-EMD (EEMD) algorithms are explored. These data-driven decompositions decompose EEG signal into intrinsic mode functions (IMFs).

The IMFs are chosen to automatically reconstruct the EEG signal. The reconstructed EEG signal contains only information correlated to a specific motor imagery task and high SNR.

The time-domain features are extracted from both the algorithms and compared for the classification of right-hand and feet MI movements. The results have been compared to determine the suitability of each method. Different classifiers, including tree, naive bayes, support vector machine, k-nearest neighbors, ensemble average, and neural network (NN), have been tested for the proposed features in order to classify the features into right hand motor imagery and feet motor imagery tasks.

Our experimental results on the BNCI Horizon 2022 dataset show that the proposed method (EEMD) with three channels outperforms > 15% with EMD based filtering with narrow NN based classification.