STAR is an R package to analyze spike trains. It provides tools to visualize spike trains and fit, test and compare models of discharge applied to actual data.
STAR works with spike train(s) from one neuron or several neurons simultaneously recorded.
STAR can perform automatic analysis of spike trains and generate html reports. The following two links show reports generated by a single command of STAR:
In the spontaneous regime. The example of function/method reportHTML.spikeTrain shows how to generate this report.
A more systematic presentation of automatic analysis results as well as of the smoothing spline based intensity estimation results (in progress) of the STAR data sets is presented in the gallery.
STAR comes with a short tutorial describing how to estimate the (conditional) intensity of a spike train. A long version of this tutorial is available from this web site together with the vignette allowing you to reproduce entirely the document (you need to have the R package snow installed in order to do that since the vignette runs computations in parallel on multicore CPUs).
If you are new to R start by looking at the appendix of the first manuscript of the "STAR trilogy" bellow. The latter is a short tutorial to R and to the basic STAR features.
Static and dynamic models for spike train analysis: Models, model diagnostics and open-source software: a manuscript describing how to estimate the conditionnal intensity of a spike train with the smoothing spline approach. Work by C Pouzat, A Chaffiol and C Gu.
Features presently implemented in STAR include:
Instantaneous firing rate estimates. They are obtained with a convolution of a Gaussian kernel with the spike train.
Test of independence of successive inter-spike intervals (ISIs) in stationary discharge regimes.
Fit of an ISI sample from a single neuron in stationary regime with various models: log-normal, inverse Gaussian, gamma, Weibull, refractory-exponential, log-logistic. The model parameters are obtained with the method of maximum likelihood and models are compared with AIC.
Test of the adequacy of the best model above.
Cross-correlograms of spike trains from different neurons recorded simultaneously.
The time rescaling (Brown et al, 2002, Neural Comp. 14: 325) / time transformation (Ogata, 1988, JASA, 83: 9) is implemented with the full collection of tests of Ogata (1988). The following example illustrates the need for the full tests collection. This is a case where Berman's test (the one used by Brown et al, 2002) is passed but not the other three (look at the bottom of the page). This does not mean that Berman's test is bad and the others good (we have example where the former fails while one of the latter is succesfull), just that more than a single test should be used. If you want to try out for yourself, the data are available (505 spike times in ASCII format, one time per line).
A new goodness of fit test is implemented based on the mapping of the "martingale": counting process - integrated conditional intensity on a Wiener Process. A proof of the mapping exists (I thank Vilmos Prokaj for telling me where to find it). The finite sample size performances of the test are excellent. See the documentation of summary.CountingProcessSamplePath in the present STAR release as well as our manuscript on goodness of fit tests for spike train models.
The statistical smoothing approach of Kass, Ventura and Cai (2003) Network: Computation in Neural Systems 14: 5. But instead of their BARS method a smoothing spline approach with package gss of Chong Gu or a GAM (General Additive Model) based approach with the R package mgcv of Simon Wood are used. We provide an R script comparing BARS and gss performances on the examples of functions: bars and barsP. Another R script shows that one can do better with smoothing splines on the example of Fig. 3 and Table 1 of Kaufman, Ventura and Kass (2005) (we do better but we still don't beat BARS on this example).
The spike-train probability model approach of Kass and Ventura (2001) Neural Comp. 13: 1713. The job is done again using the smoothing spline based approach of gss or the GAM based approach of mgcv.
Hidden Markov Models can be used in the spontaneous regime.
A demo reproducing Fig. 2-13 of Ogata (1988) shows how to work with (fully) parametric models.
Other R packages required to run STAR:
You will have to download from your favorite CRAN server and install the R package R2HTML of Eric Lecoutre in order to generate reports in html format as well package. Hidden Markov modeling requires the installation of the HiddenMarkov package of David Harte. This package is now also located on CRAN.
STAR is released under the same conditions (license, etc) as SpikeOMatic. It is installed like any other R package.
STAR comes with a full documentation and a couple of demo files as well as the vignette of the manuscript describing the basic functionalities of the software.
STAR includes several data sets:
Several neurons recorded extracellularly and simultaneously from the Cockroach antennal lobe. These data were recorded and the spike sorting was performed by Antoine Chaffiol. The five data sets include both spontaneous activity and odor responses.
Purkinje cells recording from rat cerebellar slices. Data obtained and sorted by Matthieu Delescluse.
The earthquakes data set used by Ogata (1988).
STAR is under active development any comments, suggestions and contributions are warmly welcomed.