This is an elegantly illustrated contribution that rigorously and clearly discusses the state of the art in the field of spike sorting, that is, the separation of individual units in extracellular electrophyioslogical recordings. The author discusses the problem of spike sorting, introduces the algorithms that have been proposed to solve it and provides several links to downloadable implementations.
1. - Perhaps the biggest challenge in spike sorting is that there is no clear “ground truth”. That is, as in other unsupervised clustering problems, there is no final correct answer. At least in this case, the final answer (how many neurons are near the electrode, which spikes correspond to which neurons) is very difficult to get to experimentally. It would be good to explain this problem in the discussion. - Related to this, there are a few cases where experimenters have recorded both extracellular AND intracellular data. The author cites the work of Harris et al. There is also another paper along similar lines:: M. Wehr, J. Pezaris and M. Sahani, Simultaneous paired intracellular and tetrode recordings for evaluating the performance of spike sorting algorithms. Neurocomputing, 1999. 26-7: 1061-1068.
2. Two of the most challenging problems in spike sorting are only briefly described: (i) bursting, (ii) synchronous firing. Since both bursting and synchrony are issues of high importance in neuroscience it may be worth emphasizing here that data coming from spike sorting could potentially produce biased results when answering questions about bursts or neuronal synchrony. The author could also briefly discuss the procedures that people consider to address these issues (e.g. analyzing the interspike interval distribution in the case of bursting; linear superposition in the case of synchrony).
This article provides a comprehensive and didactic overview of the basic principles and state-of-the-art in spike sorting. The collection of links to open-source sorting programs is particularly useful.
I only have a few minor comments and suggestions:
(1) It would be nice to get an idea of the volumes from which the single and multi-unit activity recorded with a standard microwire electrode can be detected and sorted (i.e., the outer and inner radius in Fig. 1). Obviously precise estimates are difficult to obtain, but the few studies that have performed simultaneous intra- and extracellular recordings as well as modeling studies constrained by these (e.g., Gold et al., 2006; Gold, Henze & Koch, J Comput Neurosci, 2007) may help to provide a rough estimate.
(2) In addition to spike detection by simple amplitude or power thresholding, some attempts have been made to use information about the spike shape already in the detection process, e.g. by using a wavelet-based detection algorithm (e.g., Hulata et al., 2002; Nenadic & Burdick, IEEE Trans Biomed Eng, 2005). Some of these approaches may be worth mentioning.
(3) The article is focused on univariate spike sorting. In many electrophysiological experiments, however, multivariate data is recorded, e.g. from tetrodes. I would be interesting to discuss the value of spike sorting for these types of multi-contact recordings.
(4) Fixed a few typos in the text.
I am very thankful to both reviewers for their positive comments. All their points where taken into consideration in the final version of the article, as detailed below.
Response to reviewer A.
1. I agree with the reviewer that the lack of ‘ground truth’ is one of the most challenging issues for spike sorting. This issue was addressed in the previous version in the section ‘Performance evaluation’. In the new version I also added some discussion of this in the conclusions. I thank the reviewer for the reference of Wehr et al, which is added in the text.
2. Bursting and overlapping spikes are two of the most challenging issues for spike sorting indeed. To discuss this, I added a section ‘Other issues’ that describes bursting, overlapping spikes as well as tetrodes and the presence of non-Gaussian clusters.
Response to reviewer B.
1. The volumes in which single- and multi-units are usually recorded are added in the legend to the figure (the size of the inner and outer circles).
2. Note that Hulata et al also used amplitude thresholding for spike detection. In the case of Nenadic and Burdick thresholding is also used with wavelet coefficients of particular frequency bands.
3. A brief description of tetrodes and how to deal with them has been added in the section ‘Other issues’.
4. Thanks for fixing the typos.