关键词: 精神分裂症, 特征提取, 特征选择, 遗传算法, 脑磁图（MEG）信号分类区分

Abstract: In order to classify the MEG signal more efficiently, an approach based on multidimensional complexity is proposed for MEG signal classification. First, several features including Autoregressive（AR） model parameters, band power, approximate entropy, Lempel-Ziv complexity are extracted from MEG signals. Then, plus-L minus-R（LRS） techniques combined with distance principle are employed to select informative channels. After channel selection, the best features are selected using Genetic Algorithm（GA）, classifiers including BP neural networks and Support Vector Machine（SVM） are used to classify the reduced feature set of the two groups. The results show that the approximate entropy and Lempel-Ziv complexity of schizophrenic are higher, it is suggested that the MEG signal is more complex. The interesting point is that most of selected channels are located in the temporal lobes, it means that the selected channels in the temporal lobes carry more discriminative information. A classification accuracy of 98.5% and 99.75% is obtained by BP neural networks and SVM respectively.

Key words: schizophrenic, feature extraction, feature selection, genetic algorithm, Magnetoencephalography（MEG） signal classification