CNN models for eye state classification using EEG with temporal ordering

Abstract

Most studies on eye states (open or closed) detection apply machine learning techniques on subject dependent eye state datasets, but subject independent data with large physiological variation between individuals has not been well explored. Temporal ordering information is important to predict eye state because EEG is a time sequence dataset. In this research, we keep the temporal ordering of the data in place. We create multiple CNN network models and select optimal filters and depth. Our CNN feature models are effective on both subject dependent and subject independent eye state EEG classifications. We got the best subject dependent result with 4 layers of CNNs with an accuracy rate of 96.51% on dataset I and 100% on dataset II. For subject s independent studies, we got the best classification accuracy of 80.47% on dataset I and we got 90.15% on dataset II.