Imaginary hand movement classification using electroencephalography

This paper proposes a method to help patients who cannot control their appendicular organs to communicate and to control devices via a binary decision by electroencephalography (EEG). We exploited 12 volunteers' EEG datasets from PhysioNet (EEG Motor Movement/Imaginary Datasets) that contain imaginary hand movement. For the signal selection, we have selected theta and alpha bands (4–15 Hz), since the signals in these bands are distinctively changed by the imagination. For the method, we have applied power spectrum density estimated by the autoregressive model (AR-model) to extract features, and then used principal component analysis (PCA) in order to reduce those features before the classification step. To measure the quality of the derived features, we used a set of classifiers including the decision tree, K-nearest neighborhood, and ensemble classifier. For the experiment, we conducted both intra-user and inter-user approaches. The leave-one-out cross validation was applied in the intra-user experiment while the five-fold cross validation was applied in the inter-user experiment. The results show that the highest average of classification accuracy is achieved by the cubic K-NN (97.08%) in the inter-user experiment and by the weighted K-NN (91.88%) in intra-user experiment.