A novel machine learning based feature selection for motor imagery EEG signal classification in Internet of medical things environment

Abstract In Internet of Medical Things (IoMT) environment, feature selection is an efficient way of identifying the most discriminant health-related features from the original feature-set. Feature selection not only finds the best informative features, but also helps in reducing the overall dimensions of the given dataset. In this paper, the actual feature-set is obtained from Brain Computer Interface (BCI) Competition-II Dataset-III motor-imagery electroencephalogram (EEG) signal using the Adaptive Auto-regressive (AAR) feature extraction technique. Based on the order (number of AR coefficients) of the AAR algorithm, two variants of datasets have been generated: 12 ( o r d e r = 6 per electrode) and 24 ( o r d e r = 12 per electrode) AAR features datasets. Here, a new fuzzified version of discernibility matrix has been proposed to determine a subset of features, which provides the best classification accuracy. In order to find the best feature subset, various types of dissimilarity measures have been used and compared with one another in our proposed fuzzy discernibility matrix (FDM) based feature selection technique. We have implemented the proposed algorithm on the given datasets using both the holdout technique as well as the 10-fold cross-validation in our study. The performances of the selected feature-subsets are evaluated based on accuracies using the Support Vector Machine (SVM) and Ensemble variants of classifiers. The empirical results obtained from our experiments in this paper is competitive in terms of accuracy and outperformed the other popular t-test, Kullback–Leibler Divergence (KLD), Bhattacharyya distance and Gini index based feature selection techniques. Our proposed FDM based feature selection algorithm using holdout technique provides 80% and 78.57% accuracies for the 12 and 24 features AAR datasets respectively. The results obtained in the holdout technique with only 50% of the best discriminant features are even better than the performances obtained while using the original feature-sets (without using any feature selection technique). Again, it gives 78.57% and 75.57% mean-accuracies from 5 × 10-fold cross-validations using only 6and 12 most discriminant AAR features from the actual 12 & 24 features-sets respectively.