Cognitive state analysis based on EEG using chaotic reservoir computing

Abstract The reliable classification and detection of mental fatigue and sleep states are crucial for cognitive state monitoring systems. In this study, it is proposed to classify pre-processed open-access electroencephalogram (EEG) data using Chaotic Reservoir Computing models. For performance comparison of the proposed model, classical Simple Multi-Layer Perceptron (MLP) and Feature-Based MLP models were tested on the same dataset. Cross-validation was performed 10 times, yielding cross-validation results of 57.51% for the Simple MLP model, 80.63% for the Feature-Based MLP model, and 81.07% for the proposed Chaotic Reservoir Computing model. It has been noted that feature extraction provides high accuracy in EEG classification in MLP models. However, the Chaotic Reservoir Computing model, in which no feature extraction was performed, demonstrated the highest cross-validation success. In terms of training and test accuracy, Simple MLP achieved 99.67%—77.27%, Feature-Based MLP achieved 94.29%—95.45%, and Chaotic Reservoir Computing achieved 90.64%—100%. As can be seen from the accuracy rates, although Simple MLP appears to have achieved very high success in training accuracy by overfitting, it showed relatively low success in the testing phase. Although the Feature-Based MLP model showed a more balanced and realistic performance in cross-validation and test accuracy, Chaotic Reservoir Computing demonstrated the highest test accuracy. The chaotic structure of the reservoir possesses the ability to effectively encode the non-linear and time-dependent dynamics of EEG signals. It is understood that the Chaotic mental fatigue method yields better results due to the chaotic nature of EEG.