An adjustment strategy on multi-session EEG data for online left/right hand imagery classification

In this research, electroencephalography (EEG) is used as an interface to communicate between patients and doctors. The signals from two electrodes (C3 and C4) are captured and used to classify Left/Right hand imagery representing YES/NO answers of the patients. In online applications, the training model mostly cannot be applied to the testing sessions due to a variation of the signals. Although some prior works employed a normalization technique, the parameters were still derived from all sessions, not just the training sessions, resulting in low prediction accuracy in real-world online systems. In this paper, we propose an adjustment strategy that can be applied online to all features by subtracting "estimated mean" and dividing "estimated interquartile rage" (IQR) or "estimated standard deviation" (SD), which are obtaining by using exponentially weighted moving average (EWMA). In our system, the features are extracted by applying the wavelet transformation, and Neural Network is chosen as our classifier. The experiment was conducted on the BCI IV data set and compared to four existing techniques: (i) non-normalized wavelet, (ii) Z-transform, (iii) CSP, and (iv) CSP with Morlet wavelet, in terms of accuracy. The results showed that our proposed method significantly outperformed the first three works and it is comparable to the last one, but ours employed the less number of electrodes.