Traditional waveform based spike sorting yields biased rate code estimates

Valérie Ventura
PNAS April 28, 2009 vol. 106 no. 17 6921–6926
Free Fulltext: http://www.pnas.org/cgi/doi/10.1073/pnas.0901771106

Much of neuroscience has to do with relating neural activity and
behavior or environment. One common measure of this relationship
is the firing rates of neurons as functions of behavioral or environmental
parameters, often called tuning functions and receptive
fields. Firing rates are estimated from the spike trains of neurons
recorded by electrodes implanted in the brain. Individual neurons’
spike trains are not typically readily available, because the signal
collected at an electrode is often a mixture of activities from different
neurons and noise. Extracting individual neurons’ spike trains
from voltage signals, which is known as spike sorting, is one of the
most important data analysis problems in neuroscience, because it
has to be undertaken prior to any analysis of neurophysiological
data in which more than one neuron is believed to be recorded
on a single electrode. All current spike-sorting methods consist
of clustering the characteristic spike waveforms of neurons. The
sequence of first spike sorting based on waveforms, then estimating
tuning functions, has long been the accepted way to proceed.
Here, we argue that the covariates that modulate tuning functions
also contain information about spike identities, and that if tuning
information is ignored for spike sorting, the resulting tuning function
estimates are biased and inconsistent, unless spikes can be
classified with perfect accuracy. This means, for example, that the
commonly used peristimulus time histogram is a biased estimate of
the firing rate of a neuron that is not perfectly isolated.We further
argue that the correct conceptual way to view the problem out is to
note that spike sorting provides information about rate estimation
and vice versa, so that the two relationships should be considered
simultaneously rather than sequentially. Indeed we show that
when spike sorting and tuning-curve estimation are performed in
parallel, unbiased estimates of tuning curves can be recovered even
from imperfectly sorted neurons.