Sparse firing frequency-based neuron spike train classification

Abstract

Peri-stimulus time histograms (PSTHs) reveal the temporal distribution of action potentials, averaged over many stimulus presentations. PSTHs have been used as model responses to solve the classification problem, in which a single response (i.e., spike train) is assigned to one of a set of response models evoked by a set of stimuli. In this study, we developed and applied a sparse firing frequency-based method to classify individual spike trains of slowly adapting pulmonary stretch receptors (SARs). Extracellularly recorded individual SAR spike trains were evoked by one of three different lung inflation volumes in anesthetized, paralyzed adult male New Zealand White rabbits. Three different PSTH-based firing frequency response models (i.e., one for each stimulus) were constructed from two-thirds of the responses to the 600 inflations presented at each volume, while the remaining one-third were used as responses to be classified. An instantaneous firing frequency representation of each remaining "test response" was computed from their individual spike trains, using one of two forms: sparse and filled. The sparse format assigned instantaneous firing rate values only in bins that contained spikes, while the filled format assigned values to intervening bins too. Classification was performed by computing the Euclidean distance between the response spike trains and the three PSTH-based models using both sparse and filled representations. When comparing the two representations with regard to classification accuracy, we found that the sparse representation does not diminish performance appreciably, while reducing computational burden significantly.