Effect of Artefact Removal Techniques on EEG Signals for Video Category Classification

Pre-processing, feature extraction, feature selection, and classification are the four submodules of a typical BCI signal-processing system. We experimented with state-of-the-art algorithms for each submodule on two datasets acquired using Emotiv EPOC and Muse headbands from 38 college-aged adults. We used two artefact-removal techniques: Stationary Wavelet Transform (SWT) denoising and an extended SWT technique (SWTSD). SWTSD improved average classification accuracy by up to 7.2% and performed better than SWT in this setting. The highest average accuracies achieved with Muse and Emotiv EPOC data were 77.7% and 66.7%, respectively. Our results show that BCI performance depends on multiple methodological choices and that appropriate signal-processing and classification methods can significantly improve results.