Modelling spike trains and extracting response latency with Bayesian binning

Abstract

The peristimulus time histogram (PSTH) and the spike density function (SDF) are commonly used in the analysis of neurophysiological data. The PSTH is usually obtained by binning spike trains, the SDF being a (Gaussian) kernel smoothed version of the PSTH. While selection of the bin width or kernel size is often relatively arbitrary there have been recent attempts to remedy this situation (Shimazaki and Shinomoto, 2007c,b,a). We further develop an exact Bayesian generative model approach to estimating PSTHs (Endres et al., 2008) and demonstrate its superiority to competing methods using data from early (LGN) and late (STSa) visual areas. We also highlight the advantages of our scheme's automatic complexity control and generation of error bars. Additionally, our approach allows extraction of excitatory and inhibitory response latency from spike trains in a principled way, both on repeated and single trial data. We show that the method can be applied to data with high background firing rates and inhibitory responses (LGN) as well as to data with low firing rate and excitatory responses (STSa). Furthermore, we demonstrate on simulated data that our latency extraction method works for a range of signal-to-noise ratios and background firing rates. While further studies are needed to examine the sensitivity of our method to, for example, gradual changes in firing rate and adaptation, the current results suggest that Bayesian binning is a powerful method for the estimation of firing rate and the extraction response latency from neuronal spike trains.