EEG is an important source of information about the activity of the brain and finds broad applications in brain-computer interface and cognitive state analysis. This study focuses on the classification of eye states using EEG signals recorded from multiple scalp electrodes. Several machine learning methods were used to differentiate between eye open and eye closed conditions using EEG recordings. The models were tested using various classification measures to have a complete measure of predictive power. The K-Nearest Neighbors model is the highest performing in terms of classification with an accuracy of 0.9609 and ROC-AUC of 0.9927. CatBoost, XGBoost, and Random Forest were also found to be good predictors by ensemble methods. To enhance the transparency of the model, explainable artificial intelligence (XAI) models, including SHAP analysis, Partial Dependence Plot (PDP) and Individual Conditional Expectation Plot (ICE) plots were used to analyze features contribution and model performance. The results of the interpretability show that the signals of certain EEG electrodes, especially in frontal and occipital brain areas are important in the process of classifying the eye state. The results indicate that machine learning with explainable algorithms can be successfully used to assist EEG-based eye state detection in addition to providing insightful information on the decision-making of the model.
Mohammad et al. (Thu,) studied this question.