A quantitative method of computer analysis of spike train data collected from behaving animals

In neurophysiological experiments in which recordings are made from behaving animals, single unit data is usually collected on-line during performance of a specific task, along with other pertinent information such as EMG. In analyzing the behaviour of a cell, the goal is to detect significant non-stationary or driven activity, either transient or sustained, and to quantify this response in terms of the latency and its variability. Different methods of analysis have been discussed at length by various authors ~,4-6. These methods include the frequently used peristimulus time histogram, and the interspike interval distribution. We have found both these approaches inadequate for reasons outlined below. Traditionally, trials are lined up in a spike raster and then summed to produce a histogram. Any consistently repeated changes in firing rate are reflected in departures of the appropriate histogram bins from the background level. The histogram provides a good visual aid, but quantitatively, latency determinations are only as accurate as the width of the bins. In addition, it is not possible to determine the variability of the individual responses from each trial. In single trials with a very well-defined response (i.e. a background firing rate of zero with a burst of spikes following the stimulus) it is relatively easy to define the onset of response and thus determine a latency for each trial. These obvious responses, however, usually comprise only a small proportion of the total population of responses obtained during recording. Thus there arises a problem of defining consistent criteria for calculating latencies for less obvious responses. A possible method of determining rate changes in the spike train of a single trial is to use the interspike interval distribution of the background period to choose thresholds for significant changes of unit firing. However, it is not accurate merely to determine the mean interval and set thresholds based on departures from this value. The interval distribution for spontaneously active, non-driven spike trains is often quite skewed (see ref. 5 for examples). So the mode (most frequently occurring interval) could be quite different than the mean interval. In addition the control period for a trial is often quite short (i.e. one or two sec). To adequately characterize the interval distribution of the control period requires a large number of intervals, many more