A Stimulus-sequence Optimization Technique And Deconvolution Method For Auditory Evoked Potentials

Auditory evoked potentials (AEPs) are series of weak electrical acitivities occurred in the internal auditory nervous system while the external sound signal transmitting in the auditory pathway. The components of AEPs with different latencies can represent the physiological conditions and responses of different parts of the nervous system under a certain stimulation sound. By detecting AEPs, the status assessment and pathological diagnosis of auditory system can be accomplished. The recording paradigm of AEP mainly consists of conventional and high stimulus rate mode. For conventional recording condition, stimulus onset asynchrony (SOA) must be adjusted with the latency range of the targeted ingredients and the short stimulation sounds (such as click sound) need to be played repeatedly with a same SOA. Then the transient AEP can be obtained by superposition averaging method. In the high stimulus rate AEP (HSR-AEP) experiments, the SOAs will become shorter than the durations of transient AEPs, which leads to an overlapping for the two successive responses to the adjacent stimuli, thus making it hard to received an unabridged transient AEP. However, a shortened recording time and aggravated stresses for auditory system can be achieved with HSR mode, which makes the HSR-AEP possess more pathological information. It was reported that the HSR-AEPs recorded under different stimulus rate could contribute to the research of adaption effect of auditory system and make it more efficient to the detection of some potential brain lesions. Therefore, the study on the separation of transient components from HSR-AEP is of important theoretical and clinical significance.From the viewpoint of signal processing, the overlapping of HSR-AEP can be modeled as s circular convolution between the transient AEP and stimulus sequence, which means that the exaction of transient response is essence a deconvolution process for the HSR-AEP. Based on this engineering model, a SOA-jittering stimulus sequence with unequal SOAs can be applied as a stimulus paradigm to implement the deconvolution operation. Currently, the deconvolution techniques used for addressing the overlap problem of HSR-AEP mainly include the maximum length sequence (MLS) deconvolution, continuous loop averaging deconvolution (CLAD), Quasi-periodic sequence deconvolution (QSD) and multi-rate steady-state averaging deconvolution (MSAD). CLAD is relatively a mature technique for recoving transient response under HSR condition. In the CLAD experiment, the stimulation sound generated based on a low-jitter binary sequence must be played in the continuous loop way to gain a sweep response as long as the sequence length by superposing average. To perform the deconvolution process easily, the time-domain circular convolution is converted into frequency-domain multiplication model, thus an inverse filter corresponding to stimulus sequence is used for deconvolving the sweep response. In the solution process, the spectum properties of stimulus sequences will largely impact on the quality of the restored signal. When spectrum values of the stimulus sequence are too small or even close to zero, the noise existed in the sweep response will be amplified to a degree resulting in a distortion. Aiming at CLAD, the principle and implementation of this techinque are introduced in detail, and on this basis, further researches on the stimulus sequence selection and calculation promotion are carried out as major works of this thesis.1. The low-jitter stimulus sequences used in CLAD are sensitive to noise. Therefore, to ensure the quality of reconstructed transient AEP, the spectrum of stimulus sequence must meet specific requirements. In order to make a sequence achieve a better noise suppression effect, a straightforward approach is to set a threshold constraint on its spectrum, which means the spectrum values must be less than a specific value in the frequence band of interest. But it is actually difficult to select a sequence that entirely meets the preset condition due to the limitation of jitter ratio (JR). To efficiently get a appropriate stimulus sequence under a low-jitter condition, a sequnence optimization strategy based on solution-space contraction differential evolution (scDE) algorithm is hereby modified to satisfy the practical needs. By changing the objective function of scDE algorithm, we can set an expected value for JR as a constraint for it and make the deconvolution gain factor dec be the optimization target. Using this new restraint method, a favorable sequence can be generated while meeting the restricting condition of JR. The optimized results are validated by reconstruction experiment, in which 6 groups of optimized stimulus sequences with JR less than 10% are used to recover AEPs and the convergence process of objective function for each preset JR condition are analyzed. The results show that the scDE algorithm with new objective function can effectively optimize stimulus sequence under low-jitter conditions, and the convergence of the objective function is considerably stable as well.2. The frequency-domain solution for CLAD technique puts a strict limitation on the spectrum properties of the stimulus sequence, making the sequence selection become a difficult problem and has inevitably brought inconvenience and restiction for this technique in the practical application. Based on the analysis of CLAD calculation model, we put forward a time-domain algorithm to accomplish deconvolution process. By constructing a sequence-based system matrix to model the convolution process, the frequency-domain solution can be solved by a linear matrix inverse filtering in the time domain. In actual use, the system matrix composed of stimulus sequence is often a large sparse matrix because of the high sampling rate. The matrix thereby may be with ill-posed problems to different degrees, leading an inaccuracy for the inverse filtering solution. In this regard, the singular value decomposition (SVD) and least square (LS) technique are used to explore the matrix property related to the bad sequence, and the regularization technique is introduced to improve the quality of reconstructed results under ill-conditioned matrix. To further verify the proposed approach,3 groups of sequences with different ill-posed problems are used to restored AEP waves under 3 different noise level. The results indicate that the time-domain solution with regularization technique could effectively deal with the ill-posedness of transformation matrix with less restriction on the sequence selection.