离散小波变换和自编码器耦合的脑电信号异常检测方法

Detection Method for Abnormal Electroencephalographic Signals Coupled with Discrete Wavelet Transform and Autoencoder

为了准确地目视检查和解释脑电图(EEG),提出了一种用于识别EEG信号中癫痫发作信号的异常检测方法。首先,使用小波变换将EEG信号分解为近似和细节系数,并根据阈值准则剔除不显著系数,以限制小波系数的数量;其次,采用自编码器对离散小波系数进行编码;然后,对EEG信号进行分析以检测异常值,通过压缩特征集进行数据重构,利用分类器从无癫痫信号中检测癫痫发作信号;最后,使用波恩大学数据库,将所提方法与既有方法进行比较。所提方法中采用了线性和非线性机器学习分类器从EEG信号中检测癫痫发作信号。实验结果表明,该方法的准确率和特异性分别达到了99.93%和100%。因此,所提方法具有良好的检测能力和鲁棒性,可以用简单的线性分类器识别EEG信号中的癫痫发作信号,适用于时间序列信号分析,同时能够检测和判断异常,也可为癫痫的诊断、治疗和评估提供客观参考,从而减轻医生的工作量,提高治疗效率。

To accurately inspect and interpret electroencephalographic(EEG),a signal anomaly detection method is proposed for identifying seizure signals within EEG recordings.First,EEG signals are decomposed into approximation and detail coefficients through the application of wavelet transform,and the number of wavelet coefficients is limited by discarding insignificant coefficients based on a threshold criterion.Second,an autoencoder is utilized to encode the discrete wavelet coefficients.Then,we analyze EEG signals to detect outliers,reconstruct data through compressed feature sets,and detect epilepsy from non-epileptic signals through a classifier.Finally,the performance of the proposed method is evaluated in comparison with established methods utilizing the University of Bonn database.Experimental results inticate that epileptic seizure signals are detected from EEG signals with classification accuracy and specificity reaching 99.93%and 100%respectively,through the employment of linear and nonlinear machine learning classifiers.The robustness and good detection capability of the method in distinguishing epileptic seizure activity within EEG signals are thus demonstrated.This approach is deemed suitable for analyzing time series signals,enabling the simultaneous detection and identification of anomalies.Therefore,it offers an objective reference for the diagnosis,treatment,and evaluation of epilepsy,potentially reducing the workload of medical professionals and enhancing the efficiency of treatment.