Spike sorting: Which clustering method should be chosen? Which circumstances affect this selection?

Abstract

There have been several researches on spike sorting, the process of detection, feature extraction and clustering neural signals generated by brain neurons. How accurate spike sorting is performed, determines how much the results are reliable. Therefore, selecting among several proposed methods for each spike sorting step is a very important task. In this paper, we want to answer this question for the clustering step. We used the most common feature extraction method, i.e. PCA and the first two principal components as features. In order to have fair judgment, 3 datasets with different noise levels were used. We compared some of the most popular methods, selecting between: model-based/non model-based, simple mixtures/Dirichlet process mixtures, Normal/t-distributions for observations, Bayesian/EM clustering. Eventually, some direct and some situation-based conclusions were obtained.