Using deep neural networks for natural saccade classification from electroencephalograms

Abstract

This paper proposes a model to classify saccades from frontal (i.e., Fp1/Fp2) electroencephalography (EEG) data into up, down, left and right directions. The aim of the model is to provide brain-computer interfaces with improved cursor control without the need for a separate eye tracking device. To test the accuracy of the model "in-the-wild," an EEG dataset with (uncontrolled) natural saccades was used, where subjects looked freely at images on a screen. EEG data was input to deep neural networks, namely a multi-layer perceptron and a convolutional neural network. As benchmarks, two systems were explored using features measured from the Fp1/2 EEG channels, as well as from the AF7/8, F7/8, FT7/8, and T7/8 channels. Experimental results show the proposed system achieving an accuracy of 72.92%, thus outperforming all benchmarks, which achieved accuracies of 51.80% and 50.72%, respectively.