波形特征的高斯混合模型锋电位分类算法

锋电位分类是进行大脑信息处理机制研究的基本步骤之一。针对锋电位信号的复杂性和非平稳性,从统计聚类的角度出发,采用高斯混合模型描述锋电位信号的概率密度函数,提出了一种新的基于波形变化率特征和高斯混合模型聚类的锋电位分类方法。首先计算锋电位的波形变化率,然后利用最大差异方法获得锋电位波形的低维特征,最后采用高斯混合模型算法对特征进行聚类,实现锋电位分类。采用开放的仿真数据分析了该算法的分类精度和可行性,然后分别利用来自5只大鼠和1只恒河猴初级视觉皮层的实测数据验证了该算法的实用性,并与主成分分析特征的高斯混合模型聚类、幅值特征的高斯混合模型聚类和变化率特征的k均值聚类等3种方法进行了比较。仿真实验中,在噪声水平为0.05、0.10、0.15、0.20时,误分率分别为1.18%±1.18%、1.41%±1.06%、2.27%±1.51%、2.98%±2.06%, 低于其他3种方法;实测实验中,恒河猴数据的J₃准则值为13.50±5.26,大鼠数据的J₃准则值为26.43±10.46。与其他3种方法相比,平均J₃准则值较大,且显著高于幅值特征的高斯混合模型聚类算法。所提出的方法表现出较高的分类精度和较好的类可分性,为实现神经元锋电位的可靠分类提供了一种有效的手段。

Spike Sorting Using Gaussian Mixture Model of Waveform Feature

Spike sorting is one of basic steps to study brain information processing mechanism. Regarding to the complexity and non-stationary of spike signals, a new spike sorting method based on waveform changing rate features and Gaussian mixture model (GMM) cluster was proposed in this work, in which the GMM described the probability density function of spike features from statistical clustering viewpoint. In the proposed method, firstly, the changing rate of spike waveforms was calculated, then low-dimensional features were obtained using maximum-difference method, and the features were finally sorted by the GMM. The accuracy and feasibility of the algorithm was measured by the simulated data from the published database, and the practicality was validated by the real data from primary visual cortex of five rats and one macaque monkey. And its performance was compared to other three sorting methods that based on principal component features and GMM cluster, based on waveform features and GMM cluster as well as based on waveform changing rate features and k-mean cluster. For the simulated data, the misclassification rate of proposed method below the other three methods, that is 1.18%±1.18%, 1.41%±1.06%, 2.27%±1.51% and 2.98%±2.06% when the noise level is 0.05, 0.10, 0.15 and 0.20, respectively. For the real data, the J₃ value of proposed method is 13.50±5.26 with monkey data and 13.50±5.26 with rat data. Compared with the other three methods, the proposed method gives the maximum J₃ value and is higher significantly than the sorting method based on waveform features and GMM cluster. It has higher precision and classification performance as well as provides an effective approach to achieve reliable spike sorting.