Vigilance Estimation Based On EEG Signals

Vigilance refers to the ability of observers maintaining their focus of attention and re-maining alert to stimuli for a prolonged period of time. Keeping vigilance above a certainlevel is very important sometimes in our daily lives. For example, when driving a car, it isvital for people to maintain vigilant. As a result, we need some methods to monitor subject'svigilance levels. Electroencephalogram (EEG) is the most commonly studied signals for vig-ilance estimation. Up to now, a lot of researchers mainly focus on using supervised learningmethods to analyze EEG data. However, it is very difficult to obtain enough labeled EEGdata and sometimes the labeled EEG data may not reliable in practice. And many labelingmethods would introduce extra noise into EEG signals.In this paper, we propose a semi-supervised clustering method to analyze EEG data andestimate vigilance states. This method makes use of a few labeled EEG data to guide featureselection and then utilizes prior knowledge of vigilance state transition to supervise EEG dataclustering. After EEG data were clustered, we carry out both spectrum analysis and phasesynchrony analysis to verify the results of estimated vigilance states got in the clusteringprocess. Experiment results of several different subjects show that our method can not onlydiscriminate between the awake state and the sleeping state through EEG signals, but alsodistinguish two middle vigilance states between those two extreme states.The contributions of this paper are as follows:Firstly, we use a semi-supervised method to analyze EEG signals. The labels for EEGsignals are very difficult to acquire. Through the semi-supervised analyze method, only alittle label information is needed to guide the clustering process.Secondly, a temporal-spatial filter is used to pre-process EEG signals. This method isvery useful for removing unrelated background signals from the original EEG signals.Thirdly, We propose a new clustering method. Compared to other traditional clusteringmethods such as normalized cut or soft cluster, our method perform better on processingEEG signals.At last, We verify our clustering results using spectrum analysis and phase synchronyanalysis.