Comparing features extractors in EEG-based cognitive fatigue detection of demanding computer tasks.

An electroencephalography (EEG)-based classification system could be used as a tool for detecting cognitive fatigue from demanding computer tasks. The most widely used feature extractor in EEG-based fatigue classification is power spectral density (PSD). This paper investigates PSD and three alternative feature extraction methods, in order to find the best feature extractor for the classification of cognitive fatigue during cognitively demanding tasks. These compared methods are power spectral entropy (PSE), wavelet, and autoregressive (AR). Bayesian neural network was selected as the classifier in this study. The results showed that the use of PSD and PSE methods provide an average accuracy of 60% for each computer task. This finding is slightly improved using the wavelet method which has an average accuracy of 61%. The AR method is the best feature extractor compared with the PSD, PSE and wavelet in this study with accuracy of 75.95% in AX-continuous performance test (AX-CPT), 75.23% in psychomotor vigilance test (PVT) and 76.02% in Stroop task (p-value <; 0.05).