One-class classification of propofol-induced sedation states using EEG signals

Brain-function monitoring is a difficult problem as the particular characteristics of the brain of an individual must be taken into account. Therefore, some particular aspects of a single brain may not be properly modeled through the use of data from other people. In addition, there are situations where brain monitoring has to happen a single time and through multiple conditions. In such a case, it is not possible to calibrate a system with previous signals from this same person as the new incoming states are unknown. Moreover, as new unseen conditions happen over time, it is not possible to create a model in relation to these incoming conditions where no information is available. For all these reasons, the use of one-class classifiers (OCC) or novelty detection techniques for brain state monitoring provides a solution for determining a shift in the condition of the brain state of an individual. In this paper, we propose to investigate the brain states in different sedation states (baseline, mild, moderate, and recovery) with electroencephalogram (EEG) signals from seven adult participants by using bandpower features in five frequency bands on 20 sensors. We propose to use the one-class nearest neighbor classifier to evaluate the detection of a change from a sedation state to posterior state. We propose to set the threshold based on the 90
^th
percentile of the current class. The results support the conclusion that an f-score of 87.56 ± 15.85% can be obtained from the baseline state to the moderate sedation state. Finally, the results indicate that the gamma band gives the best performance for discriminating between the four conditions.