An Adaptive Graph Spectral Analysis Method for Feature Extraction of an EEG Signal

The obtained electroencephalograph (EEG) signal is often contaminated by different kinds of artifacts and noises. An adaptive graph spectral analysis method is thereby proposed in this paper for feature extraction from the low signal-to-noise ratio (SNR) EEG signal with complex structures. The EEG signal is first processed by graph filter to extract the graph components of the rhythm in the graph spectrum domain. Then, the desired rhythm is extracted from the graph filtered signal using singular spectrum analysis. This method is verified by the simulated EEG signal based on a Markov process amplitude EEG model and the real EEG signal. Simulated and experimental results demonstrate that the proposed method performs better than the other extensively used methods. The proposed method can avoid components mixing for EEG signal with complex structures and extract brain rhythm from the low SNR EEG signal. Therefore, the proposed method could be potentially applied for the epileptic seizure detection.