The Detection of EEG Characteristic Waves by Using Locally Stationary Autoregressive Model

Abstract

It is much important to examine the change of waveform in short section in EEG clinical diagnosis. The goal of our research is to construct a support system for diagnosis by detecting characteristic EEG waves and labeling them in short period. The labels used in this research are hump wave, spindle and K-complex wave. They are typical characteristic waves appeared in sleep stage II. We proposed the method using the nature that generally nonstationary waveforms are able to divide into many short stationary segments. EEG was divided into many sections by using locally stationary autoregressive model and each divided section was labeled into coresponding characteristic wave by using frequency spectrum acquired by Goertzel method. We applied this method for clinical EEG data. Evaluation of this method was performed by comparing with medical doctor's labeling of hump wave. The results for applying to two normal subjects' data showed the detection rates of hump wave are over 80%. It is also suggested that our method is effective to detect not only stationary waves but nonstationary characteristic waves such as hump wave.