Comparison of Unsupervised Algorithms for On-line and Off-line Spike Sorting
32nd Annual Meeting of the Society for Neuroscience, Orlando, FL, 2002
The goal of spike sorting is to identify the extracellularly recorded multiunit activity with discrete neuronal sources. To perform this task, spike sorting algorithms consist of three independent steps: spike detection, spike projection and clustering. Attempts to accelerate these steps by using unsupervised algorithms have been made but the success of such methods is highly dependent on assumptions related to the statistics of signals relative to the noise component that is specific for the given brain area and recording technique. We present a new method that (1) does not assume spike shapes to follow any specific distribution, (2) it is unsupervised and (3) can be applied during the data collection.
First, we compare different methods of spike detection (threshold, slope, energy, template and wavelet) using the signal detection theory on simulated extracellular multiunit recordings. Second, we introduce a new method using “circular embedding” to project spike shape differences to a multidimensional space. Third, we compare the separability of such projections with that of the principal components and wavelet coefficients. Forth, we employ “superparamagnetic clustering” [1] and compare the results with K-means, and Bayesian clustering methods. Fifth, we test whether the combined method of “circular embedding” and “superparamagnetic clustering” on multi-dimensional projections is suitable to perform “on-line” during the data acquisition. [1] Blatt M, Wiseman S and Domany E (1996) Phys. Rev. Lett. 76: 3251-3254.