Current Directions in Biomedical Engineering 2015; 1:42–45

Inga Sauer*, Christopher Doerr, and Thomas Schanze

Spike sorting: the overlapping spikes challenge

Abstract: When recording action potentials (spikes) from or more neurons fire at the same time. In this case the re-
many neurons simultaneously via multichannel micro- sult is a linear superposition of the individual waveforms
electrodes the overlapping of spikes from different neu- of two or more units, in the following referred to as over-
rons is a demanding problem for detection and classifi- lapping spikes. Here, common spike sorting methods gen-
cation of spikes (spike sorting). Since multichannel elec- erally fail [3] because the superimposed spike waveform
trodes provide better possibilities to separate the superim- resembles usually none of the origin waveforms and thus
posed waveforms, we refined an algorithm for separation cannot be assigned to the correct neurons.
of overlapping spikes for the use on multichannel record-
ings and tested it on simulated data with different numbers
of signal channels and with several signal parameters. We
show that the larger the number of signal channels the bet-
ter the separation that may be achieved, especially under
demanding recording conditions.

Keywords: neural activity; extracellular recordings; multi-

channel electrodes; action potential; overlapping spikes;
source separation

DOI: 10.1515/CDBME-2015-0011

1 Introduction Figure 1: Sketch of an extracellular recording with a fibreheptode

and several neurons. Left: Heptode geometry and positions of neu-
The study of brain activity relies on the fact that most neu- rons relative to the electrode. The diameter of the heptode’s shaft
rons in the brain communicate by firing action potentials is about 100 µm. Tetrodes only have three instead of six electrodes
on the outer ring (e.g. no electrodes at positions two, four and six).
often referred to as spikes. The action potentials can be
Right: Signals recorded with the heptode. Note that the amplitude
recorded extracellularly by microelectrodes implanted in distribution of the spikes depends on the position of the neurons
the brain. It is possible to measure the action potentials of relative to the contacts of the heptode’s recording channels. Note
many neurons in a local area [1]. that almost simultaneously active neurons can lead to superim-
It is emanated that the shape of a spike arising from posed waveforms (grey) (Modified from [4]).
one neuron stays constant with time aside from special
situations like bursts [2]. Based on these assumptions it
Overlapping spikes usually only present the minor
is possible to discriminate spikes from background noise
part of the detected spikes but inappropriate analysis of
and to cluster spikes into units depending on their charac-
these overlapping waveforms may lead to wrong conclu-
teristic shapes according to the neuron they arose from, a
sions about neuronal synchrony and connectivity [5].
process often referred to as spike sorting.
The idea for the applied algorithm bases on an overlap
Common spike sorting methods are generally success-
separation algorithm developed by Michael S. Lewicki that
ful if spike waveforms are clearly distinguishable from
computes models for all temporal overlaps and all possible
background noise and from waveforms of other neurons.
neuron combinations [6].
This case occurs if the neurons fire independently from
Over the last years the use of multi-channel record-
each other. But a more demanding situation arises if two
ings became more popular because they provide a distinct
amplitude distribution for each neuron on the recording
*Corresponding Author: Inga Sauer: FB KMUB, Technische channels, also known as stereotrode effect [7]. The ampli-
Hochschule Mittelhessen (THM), Wiesenstr. 14, 35390 Giessen,
tude of a spike on a channel corresponds to the distance
Germany, E-mail: [email protected]
Christopher Doerr, Thomas Schanze: FB KMUB, Technis- between neuron and electrode tip. This effect leads to a
che Hochschule Mittelhessen (THM), Wiesenstr. 14, 35390 better spatial resolution (see Fig. 1).
Giessen, Germany, E-mails: [email protected],
[email protected]

Unauthenticated
Download Date | 6/4/16 12:22 PM
 Inga Sauer et al., Spike sorting: the overlapping spikes challenge | 43

To analyse the improvement of the separation results bution. For the testing we simulated heptodes, tetrodes,
due to the stereotrode effect we adapted the algorithm of stereotrodes and monotrodes. A sketch of the geometry
Lewicki to the use on multichannel recordings and tested of recording electrodes is shown in Figure 1, stereotrodes
it on simulated multichannel recordings with known spike consist of channels 4 and 7, monotrodes of channel 1.
times and waveforms.

2 Methods
2.1 Signal simulation

To generate data that represents the activity of neurons

as realistic as possible, we used the simulation algorithm
published in [8]. The algorithm bases on the simulation
of individual interspike interval distributions given by an
exponential distribution in combination with a Gaussian
distribution to model spike-firing rates and refractory pe-
riods. Figure 3: Block diagram of the overlapping spike separation algo-
rithm.

For a realistic incorporation of noise its amplitude dis-

tribution and spectrum were extracted from a real extracel-
lular recording. Using these noise parameters sequences
were generated and added randomly to the waveforms.
The noise recording was kindly provided by the group of
Frank Bremmer, Univ. Marburg.

2.2 Data preparation

For the application of the separation algorithm we assume
perfect detection and classification results. Thus the ap-
Figure 2: Linear superposition (red) as sum of mean spike from plication of a spike sorting algorithm can be left out and
unit 1 (blue) and mean spike from unit 2 (green) with a distinct shift the knowledge about spike times and units can be taken
against each other. The red curve is also a model for overlap separa- from the simulation results. The simulated spikes were ex-
tion.
tracted from signal by means of the known spike times. All
spikes were aligned to an overall mean spike. If an over-
During the simulation the interspike intervals are gen- lap occurs the waveform is excluded and the shifts and the
erated from these distributions and are summed up to get contained units are saved for later comparison. It is also
the spike times for each neuron. At each spike time the assumed that overlapping spikes are perfectly recognized
waveform of the firing neuron is inserted and consequently as those by a clustering algorithm. The spikes which are
in case of an overlap the voltage traces will be summed already assigned to the right units are used for model cal-
up to a new waveform. The waveforms were taken out culation within the overlap separation algorithm.
of recorded data, kindly provided by Liana Melo-Thomas,
Univ. Marburg.
Neurons were simulated as point sources and elec- 2.3 The separation algorithm
trodes as point sinks. The distance between source and
Our separation algorithm calculates models for all possi-
sink is then calculated as Euclidean distance and was used
ble neuron combinations and time shifts. But instead of
for the calculation of the characteristic amplitude distri-
using a probabilistic description to find the most appropri-

Unauthenticated
Download Date | 6/4/16 12:22 PM
44 | Inga Sauer et al., Spike sorting: the overlapping spikes challenge

calculation. The number of considered shifts s is set to the

number of samples within a spike waveform. The number
of models M is then
m
(nk )s k .
X
M= (2)
k=2

To find the model which resembles the overlapping

spike as good as possible, the cross-correlation is applied
to each model and each overlapping spike.
Seeing that the spike alignment, which is applied dur-
ing a common spike sorting, modifies the position of the
waveforms within the extracted window, we incorporate
Figure 4: Separation results for different noise amplitudes and dif-
the coefficients from shifts of ±s/2 between model and
ferent numbers of signal channels. Each bar shows the standard
error of the mean.
overlapping spike for the selection of the best fitting model
(see Fig. 3 (D and E)).

ate model as in [6] we estimated the best fitting model by

calculating cross-correlations of models and overlapping
3 Results
spikes.
The models are computed with help of the mean 3.1 Influence of noise level
spikes from each neuron. The mean spikes are shifted over
a distinct period against each other and summed up to ob- We evaluated the success of the overlapping spike separa-
tain a linear superposition (see Fig. 2). tion in percent of correctly separated overlapping spikes
The model with the highest correlation will be as- for different noise amplitudes. The number of included
signed to the observed overlapping spike. The active neu- neurons in the signal n was set to 3 as well as the maxi-
rons (units) and the spike times forming the overlapping mum number of spikes contained in an overlap model m.
spike will then be defined as the units and the spike times So for the testing conditions we set n = m. For each amount
which are included in the model which fits best. of noise 30 datasets have been analysed and as a result the
The calculation of the number of required models mean percentage value has been plotted in Figure 4.
leads to a combinatorial problem that can be described as
the number of ways to choose k overlapping units out of a
set of n units. The number of possible combinations c can
be calculated with the binomial coefficient.
m
(nk )
X
c= (1)
k=2

In Eq. 1 n is the number of units included in the signal

and k is the number of units that can be included in one
overlap thus cannot be smaller than two. If really all the-
oretically possible combinations ought to be considered it
must be m = n with m the maximum number of units that
can be contained in an overlap. All combinations of units
from two up to m have to be considered, m can be chosen
Figure 5: Separation results for different numbers of simulated
smaller than n, hence 2 <= k <= m <= n. neurons and different numbers of signal channels. Each bar shows
In the second step it must be considered that the wave- the standard error of the mean.
forms forming an overlapping spike can occur within dif-
ferent intervals. Thus the degree of overlap has to be con- Results indicate that the performance of the algorithm
sidered (see Fig. 3 (C)). increases with higher numbers of used signal channels.
Since the real spike times are usually not known and With low noise amplitudes the algorithm achieves reli-
spike alignment introduces additional temporal shifts, able results for all numbers of signal channels. With rising
several time shifts have to be introduced into the model noise level, the results of the heptode and the tetrode only

Unauthenticated
Download Date | 6/4/16 12:22 PM
 Inga Sauer et al., Spike sorting: the overlapping spikes challenge | 45

decline slowly, while the results of the monotrode fall off Nevertheless is an overlapping spike separation algo-
even with little rise of noise. The improvement even the rithm unavoidable especially when recordings containing
stereotrode achieves in comparison to the monotrode is many neurons with high firing rates have to be analysed.
clearly visible in Figure 4. The next steps will be the integration of our separa-
tion algorithm in a spike sorting algorithm which includes
3.2 Influence of the number of neurons a classification of overlapping spikes. Then it can be tested
on real recordings.
Further we evaluated the success of the separation for dif-
ferent numbers of neurons simulated in the signal. The Acknowledgment: Thanks to Frank Bremmer, University
standard deviation of noise was kept constantly at 0.03. Marburg, for providing us neuronal noise data.
The number of included neurons in the signal n varied Thanks to Liana Melo-Thomas, University Marburg,
from 3 to 6. The maximum number of spikes contained in for providing neuronal spike data.
an overlap model m was set to 3. For each number of sim-
ulated units 13 datasets have been analysed and as result Author’s Statement
the mean percentage value has been plotted in Figure 5. Conflict of interest: Authors state no conflict of interest.
The results of the monotrode decline rapidly when Material and Methods: Informed consent: Informed con-
more neurons are enclosed in the simulation, while the re- sent has been obtained from all individuals included in
sults of the heptode and the tetrode stay reasonably con- this study. Ethical approval: The research related to hu-
stant. Once again, already the use of a stereotrode leads to man use has been complied with all the relevant national
a significant improvement compared to the monotrode. regulations, institutional policies and in accordance the
tenets of the Helsinki Declaration, and has been approved
by the authors’ institutional review board or equivalent
4 Discussion committee.

We refined an overlapping spike separation algorithm

published by Lewicki in [6] for the use on multichannel
recordings and evaluated the performance on simulated References
data.
[1] Eckhorn R, Schanze T, Brosch M, Salem W, Bauer R. Stimulus-
Our results indicate that the performance increases
specific synchronizations in cat visual cortex: Multiple micro-
with rising number of signal channels especially under electrode and correlation studies from several cortical areas.
conditions with high noise amplitudes and a high number Induced rhythms in the brain. Brain Dynamics 1992: 47-80.
of neurons. Due to the fact that neurons produce spikes [2] Gibson S, Judy JW, Markovic D. Spike sorting: The first step in
with stereotypic shapes the waveforms can be quite simi- decoding the brain: The first step in decoding the brain. In:
lar. The use of multichannel electrodes leads to a greater IEEE Signal Process. Mag. 2012; 29 (1): 124–143.
[3] Pillow JW, Shlens J; Chichilnisky EJ, Simoncelli EP. A model-
dissimilarity due to the characteristic amplitude distribu-
based spike sorting algorithm for removing correlation arti-
tion and consequently provides a better separability. The facts in multi-neuron recordings. PloS one 2013; 8 (5): e62123.
runtime and the computational effort increase with ris- [4] Doerr C, Schanze T, Are heptodes better than tetrodes for
ing number of channels. However, the optimal number of spike sorting?, 9th IFAC Symposium on Biological and Medical
channels is yet unknown. Systems, 2015, 9, submitted.
[5] Vargas-Irwin C, Donoghue JP. Automated spike sorting using
The runtime increases massively with rising number
density grid contour clustering and subtractive waveform
of neurons in the signal as a consequence of the increas- decomposition. Journal of Neuroscience Methods 2007; 164
ing number of models (cf. 2). Additionally the runtime in- (1): 1–18.
creases if more neurons are simulated because of the rising [6] Lewicki MS. Bayesian modeling and classification of neural
amount of overlapping spikes. signals, Neural Comput 1994; 6: 1005–30.
Another limitation is due to the so called overfitting [7] McNaughton BL, O’Keefe J; Barnes CA. The stereotrode: A new
technique for simultaneous isolation of several single units in
problem, which occurs when an overlapping spike can
the central nervous system from multiple unit records. Journal
be represented by different models and when the model of Neuroscience Methods 1983; 8 (4): 391–397.
with the highest component number but (partially) wrong [8] Doerr C, Schanze T. The multitrode-effect influences the spike
model spikes fits best. This problem is often intensified in sorting performance: a simulation study. Biomed Tech 2014;
the case of very similar spike waves. 59 (s1):201-204.

Unauthenticated
Download Date | 6/4/16 12:22 PM