Eye movements as information markers in EEG data

Artifacts such as voluntarily and involuntarily muscle movements are usually seen as a source of noise in EEG signals. In this paper, we see artifacts as a source of information in a signal. For example, eye movements can generate a traceable change in the EEG signals. We use eye movements as an effective marker for direction of movement. We propose two experiments for classification of four eye movement directions (left, right, up and down). In the first experiment, we utilize feature partitioning method based on J48 decision tree to tackle the effect of concept drift in the training dataset resulting from dynamic non-stationarity characteristics of EEG signals. Afterward, we feed the extracted partitions to three different classifiers: multilayer perceptron (MLP) (with 10 hidden layers), logistic regression (LR) and random forest decision tree (RFDT) respectively, while comparing their classification accuracy. In the second experiment, we explored an ensemble learning mechanism as an alternative criterion to deal with the dynamic nature EEG signals. We trained the last three classifiers simultaneously on each training example, followed by a voting method to determine the dominant class label. The ensemble approach increased classification accuracy from 86.2% in the first experiment to 90.1% in the second.