基于非负矩阵分解的癫痫脑电自动检测

Automatic Seizure Detection of Electroencephalogram Signals Based on Non-Negative Matrix Factorization

针对多分类癫痫检测算法因特征维数多而导致识别率不理想的问题,提出了一种基于分数阶傅里叶变换(Fr FT:Fractional Fourier Transform)和非负矩阵分解(NMF:Non-negative Matrix Factorization)的癫痫脑电自动识别算法。首先采用Fr FT对脑电信号进行时频聚焦,并利用短时傅里叶变换(STFT:Short-Time Fourier Transform)提取脑电信号的时频特征;再应用NMF对提取的时频特征进行降维;最后将降维后的特征输入到支持向量机(SVM:Support Vector Machine)分类器中进行识别。实验结果表明,该方法能识别正常、癫痫发作间期和癫痫发作期3类脑电信号,其分类准确率可达98.8%。

In order to overcome the issue of high-dimensional features or unsatisfactory accuracy for epileptic seizure detection, we put forward an automatic seizure detection algorithm based on FrFF （Fractional Fourier Transform） and NMF （Non-negative Matrix Factorization ）. Firstly, FrFF was applied on the raw EEG （Electroencephalogram） to perform time-frequency concentration. Subsequently, STFT （Short-Time Fourier Transform） was carried out to characterize the time-frequency distribution of concentrated EEG. The generated time-frequency matrix was reshaped and then reduced by NMF. At last, SVM （Support Vector Machine ） was employed to classify extracted features. Experimental results indicate that the proposed method is capable of identifying normal, inter-ictal and epileptic EEG with an accuracy of 98.8%.