Classification of finger movements through optimal EEG channel and feature selection

Abstract

Introduction: Electrencephalography (EEG)-based brain-computer interfaces (BCIs) have become popular as EEG is accepted as the simplest and non-invasive neuroimaging modality to record the brain's electrical activity. In the current BCI research context, apart from predicting extremity movements, recent BCI studies have been interested in accurately predicting finger movements of the same hand using different pattern recognition methods over EEG data collected based on motor imagery (MI), through which a mental image of the desired action is generated when a person ideally simulates or imagines carrying out a certain motor task. Although several pattern recognition methods have already been recommended in literature, majority of the studies focusing on classifying five finger movements, were based on study designs that neglected or excluded the idle state of brain (i.e., no mental task state) during which brain does not carry out any MI task. This study design may result in an increasing number of false positives and a significant decrease in the prediction rates and classification performance. Moreover, recent studies have focused on improving prediction performance using complex feature extraction and machine learning algorithms while ignoring comprehensive EEG channels and feature investigation in the prediction of finger movements from EEGs. Therefore, the objectives of this study are threefold: (i) to develop a more viable and practical system to predict the movements of five fingers and the no mental task (NoMT) state from EEG signals (ii) to analyze the effects of the statistical-significance based feature selection method over four different feature domains (nonlinear domain, time-domain, frequency-domain and time-frequency domain) and their combinations, and (iii) to test these feature sets with different and prominent classifiers.

Methods: In this study, our major goal is not to explore the best machine algorithm performance, but to investigate the best EEG channels and features that can be used in the classification of finger movements. Hence, the comprehensive analysis of the effectiveness of EEG channels and features is performed utilizing a statistically significant feature distribution over 19 EEG channels for each feature set independently. A bulky dataset of electroencephalographic MI for EEG-based BCIs is used in this study. A total of 1102 EEG features supplied from different feature domains have been investigated. Subsequently, these features were tested with eight well-known classifiers, comprising Decision tree, Discriminant analysis, Naive Bayes, Support vector machine, k-nearest neighbor, Ensemble learning, Neural networks, and Kernel approximation.

Results: For subject-dependent analysis, the maximum accuracy of 59.17% was obtained using the EEG features that were selected the most (including (i) energy and variance of five frequency bands in frequency-domain feature set, (ii) all feature types in time-domain, time-frequency domain, and nonlinear domain feature sets) and all EEG channels by the Support vector machine algorithm. For subject-independent analysis, the maximum accuracy of 39.30% was obtained using the mostly selected EEG features (which are (i) all feature types excluding the waveform length, average amplitude change value, absolute difference in standard deviation, and slope-change value feature types in time-domain feature set, (ii) the energy and variance values of all frequency bands except gamma frequency band in frequency-domain feature set, (iii) the entropy value of five frequency bands in time-frequency-domain feature set, and (iv) SD ₂ and SD ₁/SD ₂ values where lag = 1 in nonlinear feature set) and EEG channels (which are (i) some definite EEG channels including 2nd, 3rd, 7th, 11th, 13th, 14th, and 15th channels in time-frequency-domain feature set and (ii) all EEG channels in time-domain, frequency-domain, and nonlinear feature sets) by the Support vector machine classifier.

Discussion: Experimental results demonstrate that despite the high-class number, the proposed approach obtained a modest yet considerable advancement in finger movement prediction when the results are compared to the results of similar studies. Additionally, for almost all feature sets, the statistical significance-based feature reduction method improves the prediction performance in the most of classifiers, contributing elaborate EEG channel and feature analysis. Nonetheless, in this study, we used an EEG dataset recorded from only 13 healthy subjects; therefore, a dataset covering more subjects is necessary to reach a more general conclusion.

Keywords: brain-computer interfaces (BCIs); electroencephalogram (EEG); finger movement classification; machine learning; statistical-significance based feature selection.

Figure 1

The schematic representation of the proposed methodology for classifying finger movements. EEG signals are segmented into 1-s intervals for the feature extraction process, including time, frequency, time-frequency, and nonlinear domain features. Various well-known classifiers are evaluated to distinguish BCI commands based on the extracted features. The dashed path in the “Feature Selection” block denotes an alternative analysis, where ANOVA is used to select statistically significant features that replace the full set as inputs to the classifiers.

Figure 2

Comparison of classification accuracies obtained using all features vs. ANOVA-selected features across all feature sets with the Support Vector Machine (SVM) classifier for Subject E (S4).

Figure 3

Comparison of classification accuracies obtained using all features vs. ANOVA-selected features across all feature sets with the Support Vector Machine (SVM) classifier in subject-independent analysis.