基于小波特征与动态高斯混合模型的动作电位分类研究

Spike sorting based on wavelet feature and dynamic mixture-of-Gaussians models

提出了一种结合小波时频特征提取以及动态高斯混合模型模式分类的动作电位分类新算法,以实现植入式脑电研究中非同源动作电位的非监督聚类。在阈值法检测动作电位信号的基础上,采用sym5小波变换基函数提取各个动作电位的时频特征,以提高动作电位信号在高频突变阶段的时间分辨率;考虑到动作电位信号的非平稳特性,对时频特征序列进行了分帧处理,然后分别采用高斯混合模型和贝叶斯网络模型对帧内和帧间数据进行建模,从而实现了基于动态高斯混合模型的动作电位模式分类。实验结果表明,该方法的分类性能抗干扰性及可靠性较好,仿真数据的错分率基本稳定在8.44%以内,真实数据的分类结果能较大程度贴近人工分类的结果。

To realize unsupervised spike sorting in the research of invasive brain activity,we have proposed a novel spike sorting algorithm framework based on wavelet feature and dynamic mixture-of-Gaussians clustering.After spike detection using amplitude threshold method,sym5 wavelet is employed to extract the time-frequency features representing spikes generated by different source neurons.Considering the non-stationary nature of spike train data,the wavelet time-frequency feature is divided into short time frames.Then,the dynamic clustering process proceeds in a Bayesian framework,with the source neurons modeled as Gaussian mixtures.Experimental results demonstrate that our spike sorting method achieves better robustness and reliability.Experiments on simulated spike signals show an encouraging misclassified rate below 8.44%.Furthermore,experiments on real spike signals show that the clustering results highly agree with those of human sorter.