Comparison of surrogate techniques for evaluation of spatio-temporal patterns in case of regular data

To identify active cell assemblies we developed a method to detect significant spatio-temporal spike patterns (STPs). The method, called SPADE [1,2,3], identifies repeating ms-precise spike patterns across neurons. SPADE first discretizes the spike trains in exclusive bins (defining the pattern precision, e.g. 5ms) and clips the bin content to 1 if more than 1 spike is therein. Second, STPs are mined by Frequent Itemset Mining [4], and their counts are evaluated for significance through comparison to surrogate data. The distribution of the pattern counts in the surrogate data provides p-values for determining the significance of grouped patterns. The surrogate data implement the null-hypothesis of independence, and a classical choice is to apply uniform dithering (UD) [7], i.e. independent, uniformly distributed displacement of each spike (e.g. in a range of +/- 5 times the bin width [1]). This approach does not maintain the absolute refractory period and a potentially existing ISI regularity. The binarization leads in the surrogates to a higher probability of more than 1 spike per bin, and thus by the consecutive clipping to a reduction of the spike count (up to 12%, in particular for high firing rates) as compared to the original data. This may cause false positives (cmp. [9]). Therefore, we explored further methods for surrogate generation. To not have different spike counts in the original and the surrogate data, bin-shuffling shuffles the bins after binning the original data. To keep the refractory period (RP) uniform dithering with refractory period (UD-RP) does not allow dithered spikes within a short time interval after each spike. Dithering according to the ISI distribution (ISI-D) [e.g. 7] or the Joint-ISI distribution (J-ISI-D) [5] conserves the ISI and ISI/J-ISI distributions, respectively. Spike-train shifting (ST-Shift) [6,7] moves the whole spike train, trial by trial, by a random amount, thereby only affecting the relation of spike trains to each other. Thus all of these implement different null-hypotheses, as summarized in the table below. It shows the non-/preservation (no/yes) of features in the various surrogates compared to the original data. We applied all surrogate methods (within SPADE) and compared their results using artificial, and experimental spike data simultaneously recorded in pre-/motor cortex of a macaque monkey performing a reach-to-grasp task [8]. We find that all methods besides UD lead to very similar results in terms of number of patterns and their composition. UD results in a much larger number of patterns, in particular if neurons have very high firing rates and exhibit regular spike trains. We conclude that the reduction in the spike count using UD increases the false positive rate for spike trains with CV