The Extraction Of High-Order Auditory Evoked Potential Based On Hilbert-Huang Transform

Auditory evoked potentials (AEPs) are electrical signals originated in auditory nervous system in response to external acoustic stimulation that presents to the subject repetitively. At present, auditory evoked potentials (AEPs) are successfully applied to clinical diagnosis of auditory diseases, such as evaluation of infant hearing and identifying the auditory system pathology. The conventional recording paradigm applying an equal stimulus onset asynchrony (SOA) and ensemble averaging method is perfectly used in clinic. As the stimulus rate increases to a degree that an overlapping response arises on account of the superposition of transient AEPs to the successive stimuli, this overlapped AEP is called high stimulus rate AEP (HSR-AEP), and its transient AEP can be named as high-order AEP (HO-AEP). High stimulus rates can exert strong stresses on the auditory system to be useful in assessing the auditory system adaptation as well looking into the mechanism of the generation of AEP components. In addition, HSR-AEPs also show promising in rating the depth of anesthesia or studying the effect of anesthesia, and in the study of sleep states. One also expects that using HSR paradigm to shorten AEP recording time since it will take less time when delivering the same number of stimulus events. Therefore, the study of HO-AEP is promising in clinic application.However, the overlapping problem of the HSR-AEP cannot be unwrapped by averaging method. From the engineering point of view, the HSR-AEP results from the convolution of a HO-AEP and a binary stimulus sequence. Based on this model, a SOA-jittering technique is applied---a stimulus sequence with unequal SOAs rather than conventional constant SOAs is proposed to implement the deconvolution operation. Continuous loop averaging deconvolution (CLAD) technique with lower jittering was thus developed in this consideration. However, the CLAD method that essentially uses an inverse filter would amplify the signal disturbance in specific frequency bins determined by the properties of the stimulus sequences in frequency domain. Therefore, increasing the stimulus rate does not necessarily improve the AEP recording efficiency. So many researches are focusing on enhancing the signal-to-noise-ratio (SNR) before the unwrapping process The major works of this thesis include two projects as follows:1. A nonlinear filtering technique based on Hilbert-Huang transform (HHT) has been proposed in this paper to enhance the SNR and reduce the number of averaging trials. HHT based technique consists of empirical mode decomposition (EMD) and Hilbert spectrum. EMD decomposes the EEG sweeps into a sum of intrinsic mode functions (IMFs)—different level signals with local frequencies sorted from lower to high in the time-frequency domain, frequency distribution of witch obtained from the Hilbert marginal spectrum, similar to a subband filtering procedure. Useful information or signal components and artifacts or noise with different frequency are reserved more or less separately. But the noise sharing the same frequency with AEP cannot be distinguished, so thresholding operations are then carried out to those IMFs. In this paper, we examined the energy distributions of AEP over IMF layers. Those with energies distributed mainly in 20-100Hz which are considered as containing information of interest are assigned to the genuine layer; other IMF layers with higher frequencies are considered to include primarily noise are called spurious layers; and a few other IMF layers represent very slow waves are called drifting layers. IMFs with the same frequency are similar to zero-mean Gaussian distribution, of witch the thresholds is determined by the standard deviation of corresponding IMFs. In the spurious layers, the threshold is relative low to improve the stationarity of IMFs. IMFs that classified as genuine layers contain a large amount of AEP information; however, distinct transient artifacts can also be found occasionally. Denoising in the genuine layers is mainly to remove high amplitude, very obvious artifacts or noise, for which two different thresholding approaches are proposed:Whole Rejection of the Genuine-layers (WRG) and Local rejection of the genuine-layer (LRG). Whole Rejection is accomplished by rejecting the IMF, of witch the maximum amplitude exceeds the threshold and the local rejection is accomplished by rejecting the peak that exceeds the threshold.Based on this idea, three thresholding strategies were designed. (1) Direct extraction of genuine layers IMFs to reconstruct the AEP signal; (2) IMFs in spurious-layers are filtered by corresponding thresholding, and (3) IMFs in genuine layers are processed by WRG or LRG respectively. Finally, HSR-AEP can be obtained by averaging the EEG sweep filtered by the three approaches respectively.2. Extract transient AEP using HHT combined with ensemble correlation (EC). Before applying EMD, the continuous EEG data can be segmented into EEG sweeps according to the stimulus circles. Since the EEG sweeps are induced by the same stimulus sequence, They contain the same AEP components contaminated by strong background noise. After the EMD, the SNR of corresponding IMFs vary, where different correlation among the IMFs exists, especially for those IMFs with higher SNRs. Therefore, the EC functions can be regarded as weighting filters applying on the corresponding IMFs. To further enhance such correlation, we attempted to increase the SNR of EEG sweeps before conducting EMD by sub-group averaging. In comparison with thresholding approach, this method is able to enhance the signals in IMFs with less subjective parameters. The testing results on the same data set demonstrate that the noise can be effective suppressed without requiring a prior information of the signal, and artificial intervention, which will promote the practical application in the future.