Nonlinear Dynamic Extraction Of Evoked Potential

Dynamic evoked potential (EP) has proved to be valuable to the research of neural systems. Like many biological signals, the dynamic extraction of EP is corrupted by noise from background ongoing spontaneous electroencephalogram (EEG). Since the amplitude of this noise is much higher than that of the EP response, the signal to noise ratio (SNR) is typically low. As the spectra of EP and EEG significantly overlap, ordinary bandpass filtering methods fail to properly extract the EP component. Furthermore, EP originates from the complex neural systems and there may be abundant and important information during its dynamic behavior, so the averaged EP can't satisfy the demand of clinical and medical research. As a result, the dynamic extraction of nonlinear EP has been a focus and also a difficult point in these recent years.The main purpose of this thesis is to try to extract dynamically the simulated nonlinear brainstem auditory evoked potential (BAEP) and event-related potential (ERP) from background noise. Two parts of research are included. First, the simulated nonlinear BAEP/ERP is constructed and dynamic extraction is carried out. Second, the method used by simulated BAEP is tested by clinical BAEP recordings.The main methods of this thesis are as follows. Nonlinear BAEP/ERPtime-series is made with the initial phase shifted randomly according to biological range, and then white noise with different amplitude is added to the time-series to get 4 models with different SNR. Radial basis function neural network (RBFNN) is chosen to carry out the dynamic extraction as a dominant method. RBFNN is a special type of neural network linear-in-weight in nature and having nonlinearprocessing properties. Typically it can achieve better approximation performance for a higher SNR, which unfortunately is not the case of our simulated signal (0 db, -5dB and -10dB respectively). Moving window average (MWA), an approach which can raise SNR through less trials of averaging while at the same time can retain the dynamics of time-series, is employed as a pre-processing technique. Singular value decomposition (SVD) is also used as a denoising method here. SVD follows MWA to get rid of the residue noise in the pre-processed time-series. Then RBFNN is applicable to do the following dynamic extraction work.The results are as follows. For the two simulated BAEP whose initial phase shift is set to 1% of its duration, (1) if the original SNR is -5dB, the relative mean square error (RMSE) of the extracted signal fluctuates by 4.23%; (2) if the original SNR is as low as -10dB, the RMSE is 6.68% or so. As for the case of two simulated ERP with a SNR set to be 0dB, (1) if its initial phase is shifted during 1% of its duration, a RMSE of 1.18% can be obtained; (2) else if initial phase is during 5%, the RMSE is 11.56%. Furthermore, if this method is applied to clinical BAEP, its also gives better results than averaging.It can be concluded that since the simulated nonlinear EP models provide suitable tools for the dynamic extraction of nonlinear EP, the integration of MWA, SVD and RBFNN can achieve better performance even for a relatively low SNR. The method of this thesis can be extended to the dynamic extraction of other kinds of EP.