Foundations of Neuronal Representations - Brain-Machine Interface Engineering - page 27

of interpreting this population activity is to determine information flow in networks of neurons
that could be serially connected or excited through diverging or converging chains. Sherrington
[ 9 ] introduced the concept of neural ensemble as a population of neurons involved in a particular
computation in a distributed manner (each neuron shares a part of the processing). Because cortical
neurons have large overlaps both in the extracellular potential generated by axonal projections and
within dendritic trees, BMI implementations will need to interpret the many-to-one or one-to-
many mappings. Moreover, the neural firing is not a deterministic process due to the tiny electrical
fluctuations in the membrane produced by the other neurons, and the fact that the neuron normally
operates just below threshold, waiting to fire as will be described below.
This microscopic scale of analysis can provide insight to the communication (both rate and
time codes) of ensembles of local distant neurons. The variability in extracellular voltage potentials
that produce variations in the neuronal recordings are attributed to cell morphology and the resis-
tivity of the surrounding tissue. For cellular environments, the maximum measured potential has
been shown to decrease in a spherical volume with a radius of 2 mm as measured from the soma
of the pyramidal cell [ 22-24 ]. Experimentally and in simulation, the amplitude of the extracellular
field drops off sharply at distances of 50-100 µm from the cell body [ 22 ]. Experimental access to the
raw extracellular potentials enables analysis of the synaptic dendritic activity as LFPs (0.5-200 Hz)
as well as spikes (300-6 kHz) originating from the cell body and axon. Depending upon the
application, postprocessing of extracellular potentials is enhanced by low- or band-pass filtering.
Ultimately, in single-unit neural analysis, it is the action potential that is of great interest to neural
ensemble neurophysiologists. Analysis of action potentials provides a consistent waveform sig-
nature that can be tracked, which is rooted in the fundamental elements of the nervous system.
Therefore, it is critical to detect peaks in the potential above the noise floor from local and distant
neurons that must be assessed directly from the raw potentials. Electrophysiological parameters of
the action potential (amplitude and width) are then used to identify whether a recording electrode
contains any spurious signals (e.g., electrical noise, movement artifact). In a preprocessing proce-
dure called spike sorting, the peak-to-peak amplitude, waveform shape, and interspike interval
can be evaluated to ensure that the recorded neurons have a characteristic and distinct waveform
shape when compared with other neuronal waveforms in the same and other recording electrodes.
The goal of spike sorting is to reliably and confidently select action potentials from single neu-
rons and extract the firing times. Hence, a variety of sorting techniques have been proposed [ 18 ],
with the most common approaches being template matching and principal component analysis
(PCA)-based analysis of waveform variance. Extensions of the basic techniques have been pro-
posed using automated, manual, and multielectrode implementations [ 20 , 25-28 ]. Despite the
great advances in the spike sorting technique, there is often much variability in the analysis due to
the experimenter [ 21 ].