Recognition of Upper-limb Movement Using Electroencephalogram Signals with Deep Learning

The recognition of electroencephalogram (EEG) based motion intention is usually performed using features that are extracted from different frequency bands including delta, theta, alpha, beta, and gamma waves. However, such an approach of feature extraction requires rigorous analysis and validation to develop an efficient machine-learning algorithm. The aim of this study is to classify rest and hand movements using the raw EEG signals and deep learning (DL) network. Healthy adult subjects age ranging 18-25 years performed hand movement four times each. The DL network was trained using subject-specifically with window sizes varying between 300ms and 1500ms. In the post-processing step, the majority voting scheme was employed. The results showed that the DL algorithm can predict the hand movement with more than 80% accuracy using 600ms or greater window size. The prediction accuracy is improved to more than 90% using the majority voting scheme. However, the window size of greater than 600ms did not significantly improve the accuracy values. When comparing between 300ms and 600ms window sizes, the accuracy increases from 77.5% to 82.7% (p<0.05). Such improvement was also reflected in both ROC curves and ranking analysis, where 600ms window size yielded higher AUC values and scores compared to the 300ms window size. The results of this study is valuable in human-computer interaction as well as in controlling devices including prosthesis, exoskeleton, virtual objects, computer typing, and wheelchair.