Diagonal loading common spatial patterns with Pearson correlation coefficient based feature selection for efficient motor imagery classification.

Abstract

The conversion of a person's intentions into device commands through the use of brain-computer interface (BCI) is a feasible communication method for individuals with nervous system disorders. While common spatial pattern (CSP) is commonly used for feature extraction in BCIs, it has limitations. It is known for its susceptibility to noise and tendency to overfit. Moreover, high-dimensional, and irrelevant features can make it harder for a classifier to learn effectively. To address these challenges, exploring potential solutions is crucial. This paper introduces Regularized CSP with diagonal loading (DL-CSP) and Pearson correlation coefficient (PCC) based feature selection to extract the most discriminative motor imagery EEG (MI-EEG) features. Three classifiers in an ensemble are considered; bidirectional long short-term memory (Bi-LSTM), K-nearest neighbors (KNN) and naïve Bayes (NB). Decision level fusion through majority voting is exploited to leverage diverse perspectives and increase the overall system robustness. Experiments have been implemented using three publicly available datasets for MI classification; BCI competition IV-IIA (data-1), BCI Competition III-IVa (data-2), and a stroke patients' dataset (data-3). The accuracy achieved, according to the results, is 86.96% for data-1, 91.70% for data-2, and 85.75% for data-3. These percentages outperform the accuracy achieved by any state-of-the-art techniques.