Ventura Valérie
Department of Statistics and the Center for the Neural Basis of Cognition, Carnegie Mellon University, Pittsburgh, PA 15213, USA.
Neural Comput. 2009 Sep;21(9):2466-501. doi: 10.1162/neco.2009.12-07-669.
Current spike sorting methods focus on clustering neurons' characteristic spike waveforms. The resulting spike-sorted data are typically used to estimate how covariates of interest modulate the firing rates of neurons. However, when these covariates do modulate the firing rates, they provide information about spikes' identities, which thus far have been ignored for the purpose of spike sorting. This letter describes a novel approach to spike sorting, which incorporates both waveform information and tuning information obtained from the modulation of firing rates. Because it efficiently uses all the available information, this spike sorter yields lower spike misclassification rates than traditional automatic spike sorters. This theoretical result is verified empirically on several examples. The proposed method does not require additional assumptions; only its implementation is different. It essentially consists of performing spike sorting and tuning estimation simultaneously rather than sequentially, as is currently done. We used an expectation-maximization maximum likelihood algorithm to implement the new spike sorter. We present the general form of this algorithm and provide a detailed implementable version under the assumptions that neurons are independent and spike according to Poisson processes. Finally, we uncover a systematic flaw of spike sorting based on waveform information only.
当前的尖峰分类方法专注于对神经元的特征尖峰波形进行聚类。所得的尖峰分类数据通常用于估计感兴趣的协变量如何调节神经元的放电率。然而，当这些协变量确实调节放电率时，它们会提供有关尖峰身份的信息，而到目前为止，为了进行尖峰分类，这些信息一直被忽略。本文描述了一种新的尖峰分类方法，该方法结合了波形信息和从放电率调制中获得的调谐信息。由于它有效地利用了所有可用信息，因此这种尖峰分类器产生的尖峰误分类率比传统的自动尖峰分类器更低。这一理论结果在几个例子上得到了实证验证。所提出的方法不需要额外的假设；只是其实现方式不同。它本质上是同时而不是像目前这样顺序地执行尖峰分类和调谐估计。我们使用期望最大化最大似然算法来实现新的尖峰分类器。我们给出了该算法的一般形式，并在神经元是独立的且根据泊松过程产生尖峰的假设下提供了一个详细的可实现版本。最后，我们揭示了仅基于波形信息的尖峰分类的一个系统性缺陷。
