A non-parametric method for automatic neural spike clustering based on the non-uniform distribution of the data

Abstract

In this paper, we propose a simple and straightforward algorithm for neural spike sorting. The algorithm is based on the observation that the distribution of a neural signal largely deviates from the uniform distribution and is rather unimodal. The detected spikes to be sorted are first processed with some feature extraction technique, such as PCA, and then represented in a space with reduced dimension by keeping only a few most important features. The resulting space is next filtered in order to emphasis the differences between the centers and the borders of the clusters. Using some prior knowledge on the lowest level activity of a neuron, such as e.g. the minimal firing rate, we find the number of clusters and the center of each cluster. The spikes are then sorted using a simple greedy algorithm which grabs the nearest neighbors. We have tested the proposed algorithm on real extracellular recordings and used the simultaneous intracellular recordings to verify the results of the sorting. The results suggest that the algorithm is robust and reliable and it compares favorably with the state-of-the-art approaches. The proposed algorithm tends to be conservative, it is simple to implement and is thus suitable for both research and clinical applications as an interesting alternative to the more sophisticated approaches.

Deflation technique for neural spike sorting in multi-channel recordings

“…Spikes that come from the neuron recorded intracellularly are labeled the with the red stars. Results of sorting algorithm from [15] are given in the third column, where the black dots mark the detected centers of the clusters.…”

“…Only to demonstrate that spike sorting is very difficult for such recording we apply directly the method based on a feature extraction, described in [15]. First, we detect the spikes from each recording site using a simple spike detection method based on filtering of the recorded signal with a specially designed Volterra filter that emphasis spike shaped waveforms [5], [16].…”

“…The plots of the features are shown next to the corresponding signals, in the second column on figure 1. In the third column we give the outputs of the algorithm from [15]. The spikes that are left unsorted by the algorithm are automatically eliminated and not plotted.…”

“…We also propose a clustering algorithm (see [15] for details) that is applied within each iteration of the internal loop. As this clustering algorithm is specifically designed for neural recordings, it presents an improved robustness compared to classical approaches in this specific application.…”

“…For classifying the spikes, we use a simple approach which also relies on the non-uniform distribution of the data: each cluster is initialized by its singleton center and the clustering is performed by classifying one spike at a time, in such a way that if an element belongs to a cluster then so does its nearest neighbor. This procedure leads to a very robust solution as shown in [15].…”