A Comparison of Time, Frequency and ICA Based Features and Five Classifiers for Wrist Movement Classification in EEG Signals

This study presents a comparison of two methods to extract features for the classification of wrist movements (flexion, extension, pronation, supination). For the first method, a set of 160 features was extracted from the filtered time and frequency domain EEG data and its alpha, beta, and theta bands. For the second method, a set of 40 features per movement type was extracted from the ICA-calculated source signals. The value of the Davies-Bouldin cluster separation index for each feature was used for selecting the best five features from each set so as to avoid the subjective selection or rejection of any of the features. Finally, five different kinds of classifiers were chosen to obtain classification error rates with which to compare both techniques. The results showed the advantage of using ICA source signals for wrist movement classification purposes, at least as compared to the simple time and frequency domain features. Left and right movements were correctly identified with accuracies ranging from 70% to 96%. However, the methodology presented here did not succeed in distinguishing the subclasses (e.g., flexion versus extension) with accuracy above 70%. This suggests that additional work is needed to explore different features as well as classifiers.