1D Convolutional Neural Network-Based Hierarchical Classification of Eye Movements Using Noncontact Electrooculography

This study addresses the discomfort and challenges posed by traditional electrooculography (EOG) measurement methods that require skin-contact electrodes by developing a non-contact EOG signal measurement device. The primary objective of this research is to implement a hierarchical deep learning model for classifying data collected from a non-contact EOG device into five types of eye movements. In this study, indium tin oxide (ITO) film was employed as a capacitive coupled electrode to measure EOG signals without skin contact. A hierarchical classification model based on a 1D convolutional neural network (CNN) was proposed for signal analysis, and K-fold cross-validation was used to train and validate the model. Unlike conventional Ag/AgCl electrodes, the proposed device enables EOG signal measurement without direct skin contact. ITO film was integrated into glasses to create non-contact electrodes while minimizing signal noise. Additionally, various signal feature extraction methods, including Fast Fourier Transform (FFT), Band Power, and Hilbert-Huang Transform, were applied to enhance classification accuracy. The model using the FFT method achieved 73% accuracy in Step 1 classification, with 84% accuracy for vertical channels and 81% for horizontal channels in Step 2. The Band Power method yielded 59% accuracy in Step 1, with 62% accuracy for vertical channels and 90% for horizontal channels in Step 2. The Hilbert-Huang Transform method produced 68% accuracy in Step 1, with 63% for vertical channels and 66% for horizontal channels in Step 2. The proposed non-contact EOG measurement system demonstrated improved usability and performance over traditional methods. It is expected to achieve practical applications in human-computer interaction (HCI) systems for patients with neurological disorders.