Detecting attention in Hilbert-transformed EEG brain signals from simple-reaction and choice-reaction cognitive tasks

The aim of this study was to investigate supervised machine learning approaches for detecting attentive brain states in the electroencephalogram (EEG) signal. EEG was recorded during methodologically similar tasks with different attentional loads: choice-reaction task (CRT) and simple-reaction task (SRT). This approach minimalizes the influence of other cognitive processes or motor preparation on classification results and thus shows the real discrimination of attentive states. We applied a Hilbert transformation to single trial EEG data to extract selected signal features and then compared the effectiveness of three classifiers: Extra Trees (ET), Support vector machines (SVM) and logistic regression; as well as two methods of feature selection: an ANOVA-based method and Sequential backward floating selection (SBFS). ET and SVM classifiers and logistic regression yielded similar classification results. Classification accuracy was up to 100% for individual subjects and 89% was the average classification accuracy for all subjects after SBFS with the use of ET and logistic regression. ET achieved the highest precision (91%) and specificity (91 %), whereas highest sensitivity (89%) was observed for LR.