International Journal of Advanced Science and Technology
Vol. 3, February, 2009
Experimental Study on Neuronal Spike Sorting Methods
Jianhua Dai1,2 , Xiaochun Liu1,2, Huaijian Zhang1,3,4 , Yu Yi1,3,4, Jingjing Wang1,3,4,
Shaomin Zhang1,3,4, and Xiaoxiang Zheng 1,3,4
1
Qiushi Academy for Advanced Studies, Zhejiang University, Hangzhou 310027,
P.R. China
2
College of Computer Science and Technology, Zhejiang University,
Hangzhou 310027, P.R. China
3
College of Biomedical Engineering and Instrument Science,
Zhejiang University, Hangzhou 310027, P.R. China
4
Key Laboratory of Biomedical Engineering of Ministry of Education,
Zhejiang University, Hangzhou 310027, P.R. China
jhdai@zju.edu.cn
Abstract
When recording extracellular neural activity, it is often necessary to distinguish action
potentials arising from distinct cells near the electrode tip, a process commonly referred to as
spike sorting or action potential sorting. Sorting of neuronal spikes plays a very important
role in coding of neural information, which is a prerequisite for studying the brain function.
In this paper, five major action potential classification methods including Template Matching,
Wavelet Transform, Principal Component Analysis, Back-Propagation (BP) Neural Network,
Two-stage Radius Basis Function Network are studied. Under the conditions of different
levels of background noise, the performances of these methods are tested. This work may be
helpful to choose classification method.
1. Introduction
Most neurons in the brain transmit information by firing action potentials [1-2]. These
time-voltage action potentials can be recorded with a microelectrode, which can often acquire
the signals of many neurons in a brain region [3]. The neurophysiologist wish to make the
functions of each neuron understood, therefore, discriminating these signals from the others is
the first step, and also critical step [4-5]. Spike sorting is the process of detecting action
potentials from extracellular signals and assigning them to individual neurons. Early
development aims at assisting researchers in studying brain functions off-line. Recent
applications include brain-computer interfaces and neural prostheses for people suffering
from nervous system traumas. These efforts all build on the capability of accurate automatic
decoding of neuronal signals, which imposes a statistical computing task replete with open
problems [2]
In many cases, this work can be accomplished with a simple threshold method. Often,
however, just measuring the voltage of neurons to do classifying is a challenge due to a high
amount of background noise, and also because neurons in a local area often have action
potentials of similar shape and size. Therefore, it is necessary to do further research in spikeclassifying algorithms.
41
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
2. Methodology
2.1. Second-order headings
Figure1. a)Waveforms of the Raw Signal from the Simulator; b). Specifications of the
Action Potential from the Simulator
The signals used in the paper are from the Neural Signal Simulator. The raw signal is in
Fig.1.a. The Neural Signal Simulator produces three different shaped action potentials on
each of 128 output channels(Fig.1.b). The circuit simulates the output of a Cyberkinetics 100
electrode array. Each virtual electrode is simulated as detecting microvolt signals from
three separate neurons located at different distances from the recording site. The amplitudes
of the three action potentials as well as the kinetics of their responses differ in a manner
consistent with real world signals.
The simulation on each output channel consists of a sequence of three individual action
potentials that ‘fire’ one after the other at a 1s intervals. This firing sequence repeats nine
times. Then, every 10 seconds, a one-second burst of activity is simulated. The burst consists
of the same train of three individual action potentials, but they are repeated with an inter
action potential interval of 10 milliseconds (Fig.1.b). To achieve the results of the method of
detection and classification in a range of different degrees of the noise in the environment, we
construct the noised data by adding white Gaussian noise to the raw signal in the
measurement of dB that is the scalar of SNR which specifies the signal-to-noise ratio(Fig.2).
The background noise is added using the function Awgn() in Matlab. This function adds
white Gaussian noise to the vector signal. The scalar SNR specifies the signal-to-noise ratio
in decibels. This syntax assumes that the power of the vector signal is 0 dB.
2.2. Template Matching
In order to recognize an object, we compare it to action potentials of the similar objects
that we have stored in memory. By comparing with variety of stored candidates, we can
identify the object by the one that it most closely resembles [6-11].
2.3. Wavelet Transform
There is a generally observation that the differences between action potentials primarily
come to transient differences in high frequency features (like sharp edges and steep leading or
trailing slopes) and/or in low frequency features (like the duration of the re-polarization
phase). Thus in this paper, we adopt the Wavelet based Spike Classifier (WSC) method,
42
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
where it use the quantification of energy found in specific frequency bands at specific time
locations during each action potential profile to classify different waveforms[12-14].
Figure2. Signals Added White Gaussian Noise in Different SNR
2.4. Principal Component Analysis
Principal Component Analysis (PCA) has been widely used in feature extraction from
complex and high dimensional data in many fields, such as signal processing, image
processing and pattern recognition. It reduces the dimensionality of the feature space by
creating new features that are linear combinations of the original features [15-18].
2.5. K-means Cluster
K means clustering is a simple unsupervised learning algorithm. It can be divided into 4
steps. a).Initialize the cluster center. b).Update the cluster that every sample belongs to.
c).Update the centers of every cluster. d).If the termination condition is reached, terminate the
iteration, else go to step b)[19-20].
2.6. Back-Propagation (BP) Neural Network
BP neural networks uses Back Propagation algorithm to learn the weights. The back
propagation algorithm is as follows: a). Forward propagate the input from input layer to
output layer; b). Back propagate the error from output layer to input layer. Update the
weights matrix between input layer and hidden layer, and the weights matrix between hidden
layer and output layer; c).If the termination condition is met, the algorithm ends, else returns
to a)[21-23].
2.7. Two-stage Radius Basis Function Network
43
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
There are three layers in typical RBF networks, input layer, hidden layer and output layer.
Input layer is made up of conception units, which receive input from outside. Hidden layer
applies a nonlinear transformation between input layer and hidden layer. Output layer
calculates the linear weighted sum of hidden units’ output and provides a result after a linear
transformation [24-26].
3. Experiment Design
Original spike signals for sorting are from Neural Signal Simulator as mentioned in part A.
signal source. As said before, there are three types of spikes produced from the simulator,
with high SNR. Hence we can detect every spike exactly and construct a group of action
potentials for classifying. For those original action potentials (fig), we can easily utilize these
five methods to cluster them into 3 types, as the standard to evaluate the correct rate of other
classification methods later. After that, white noise is added to original clean spikes to create
different SNR signals (0, -10, -15, --20, -25, -30, -35) to check the effectiveness of these
methods on spike sorting.
4. Results
Applying the five methods to classify simulated action potentials in different SNR, we get
the correct ratio of these five methods in Table 1 and Fig.3.
Table 1. Correct ratio of these five methods in different SNR
SNR(-dB)
5
44
Template Matching
1
PCA
1
RBF
1
BP
1
Wavelet Transform
1
21
1.0000
0.9767
0.9900
0.9810
1.0000
22
1.0000
0.9687
0.9830
0.9470
0.9987
23
1.0000
0.9673
0.9820
0.9660
0.9973
24
1.0000
0.9653
0.9720
0.9410
0.9947
25
0.9987
0.9573
0.9640
0.9050
0.9873
26
1.0000
0.9493
0.9630
0.9080
0.9861
27
1.0000
0.9453
0.9610
0.8790
0.9773
28
0.9973
0.9313
0.9610
0.8520
0.9653
29
0.9853
0.9407
0.9230
0.8500
0.9487
30
0.9873
0.9247
0.9050
0.8090
0.9367
31
0.9747
0.9107
0.8670
0.7930
0.9127
32
0.9700
0.8907
0.8320
0.7540
0.8933
33
0.9500
0.8707
0.8050
0.7150
0.8680
34
0.9207
0.8513
0.7760
0.6180
0.8313
35
0.8913
0.8253
0.7390
0.6210
0.7993
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
36
0.8713
0.7927
0.7030
0.6070
0.7627
37
0.8420
0.7433
0.7120
0.5940
0.7387
38
0.8473
0.6833
0.6870
0.5540
0.6887
39
0.7707
0.7007
0.6540
0.5490
0.6527
40
0.7293
0.6413
0.6260
0.5020
0.6080
41
0.6606
0.6013
0.5850
0.5090
0.5933
From the Table 1 and Fig 3, we find that all methods can accurately classify the three
action potentials. However, as the lower of the SNR, these performances are much different.
The correct ratio of template matching method has no significant change until the SNR is 30dB, and it is the best in the five methods, and BP is the worst. When the SNR decreases to
a certain degree, less than -30dB, all these methods cannot competently classify the three
action potentials accurately. And the waveforms of action potentials in the noised signal have
become chaos on the whole.
Fig.3. Correct ratio comparison of these methods in different SNR
5. Discussion
Firstly, we make comparisons between RBF network and BP network methods used
in this article. The result shows that RBF network performed better than BP network in
action potentials classification. The two-stage RBF network has short training time, but
RBF centers selected by K-means method have great influence on its performance.
Because the amount of data is small in our experiment, the RBF centers account for a
45
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
relatively larger proportion of the total data. Thus, the centers can easily reflect the
distribution of the input samples, which helps to improve the correct ratio of
discrimination. In this paper, the initial centers are selected by ‘farthest points’ method
instead of being randomly selected. So the RBF centers can reflect the distribution of
the samples better, which makes the correct ratio of discrimination with RBF networks
relatively stable.
Secondly, we make comparisons between DWT and PCA methods used in this
article. Since the amplitude and peak are key features in discriminating action
potentials, while the shape of waveform is not easy to quantify. Therefore, using S.D.
of DWT coefficients as feature extracting vectors, takes well advantage of wavelet
properties. Compared with widely used PCA method, we got higher correct rate in
sorting and have more specified physiological explanation. More importantly, DWT
served as both filtering and classification functions in the processing, because it will
not result in shape distortion of the original signal.
From all the methods, template-matching methods yielded the best classification
accuracy compared to spike-shape features, principal components, and other methods.
Moreover, it can classify the action potentials online.
6. Conclusion
Sorting of neuronal spikes plays a very important role in coding of neural
information, which is a prerequisite for studying the brain function. There are still many
problems that limit the robustness of many of the current methods, such as nonstationary background noise, electrode drift and proper spike alignment. Possibly the
most restrictive assumption of most methods is the assumption of stationary spike
shapes. And currently there are no methods that can accurately classify highly
overlapping groups of bursting action potentials. Decomposing overlapping action
potentials with non-stationary shapes is largely an unsolved problem. Techniques that
use multiple electrodes and incorporate both action potential shape and action potentials
timing information are promising in surmounting this problem.
Acknowledgements
The work has been supported by the National Natural Science Foundation of China (No.
60703038) ,the Excellent Young Teachers Program of Zhejiang University and the
Research Foundation of Center for the Study of Language and Cognition of Zhejiang
University.
References
1. Rieke F., Warland D., Stevenck D. R. V., et al. Spike:Exploring the neural code. Cambridge, MA: The MIT
Press (1996)
2. Michael S.: A review of methods for spike sorting: the detection and classification of neural action potentials.
Network: Comput Neural Syst. 9, R53-R78 (1998)
3. Welsh J. P., Schwarz C.: Multi-electrode recording from the cerebellum. Methods for Neural Ensemble
Recordings, 5;79-100(1999)
4. Schmidt E M. :Computer separation of multi-unit neuroelectric data: a review. Neurosci. Meth, 18: 192195(1984)
46
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
5.
Lewick M.S.: A review of methods for spike sorting: the detection end classification of neural action
potentials. Network: Computation Neural System, 9: 53-78(1998)
6. Zouridakis G., Tam D.C.: Identification of reliable spike templates in multi-unit extracellular recording using
fuzzy clustering, Computer Methods and Programs in Biomedicine, 61(2):91-98; (2000)
7. Zhang P.M., Wu J.Y., Zhou Y, et al,: Spike sorting based on automatic template reconstruction with a partial
solution to the overlapping problem .J. Journal of Neuroscience Methods, 135:55-65( 2004)
8. Jansen R.F., Temaat A.: Automatic waveform classification of extracellular multi-neuron recordings. J.
Journal of Neuroscience Methods, 42: 123-l32(1992)
10. Chandra R, Otican L.M.: Detection, classification, and superposition resolution of action potentials in
multiunit single-channel recordings by on-line real-time neural network. J.IEEE Trans.Biomed Eng.,
44(5): 403-412(1997)
11. Lewicki M.S.: Bayesian modeling and classification of neural signals. J. Neural Computate, 6: 1005-1030
(1994)
12. Zouridakis G, Tam D.C.: Multiunit spike discrimination using wavelet transforms. J; Computers in Biology
and Medicine, 27(1): 9-18(1997)
13. Letelier J.C., Weber P.P.: Spike sorting based on discrete wavelet transform coefficients. J; Journal of
Neuroscience Methods, 101(2): 93-106(2000)
14. Glaser E.M. : Separation of neuronal activity by waveform analysis. Advances in Biomedical Engineering vol
1, pp 77-136 (1971)
15. Glaser E. M., Marks W. B. : On-line separation of interleaved neuronal pulse sequences. Data Acquisition
Process. Biol. Med. 5 137-56 (1966)
16. Feldman J. F. , Roberge F. A.: Computer detection and analysis of neuronal spike sequences. Inform. 9, 185-97
(1971)
17. Wheeler B. C. , Heetderks W.J.: A comparison of techniques for classification of multiple neural signals. IEEE
Trans. Biomed. Eng. 29 752-9(1982)
18. McNaughton B. L. , O’Keefe J. , Barnes C.A. : The stereotrode: a new technique for simultaneous isolation of
several single units in the central nervous system from multiple unit records . J. Neurosci. Meth. 8, 391-7(1983)
19. Mantao Xu, Franti: A heuristic K-means clustering algorithm by kernel PCA. Image Processing, Page(s):3503
- 3506 Vol. 5 (2004)
20. Wilkin, Gregory A., Huang Xiuzhen: K-Means clustering algorithms: implementation and comparison.
Page(s):133 -136(2007)
21. Tom M.: Machine Learning, China Machine Press, May(2007)
22.Khashayar Mirfakhraei, Kenneth Horch: Classification of Action Potentials in Multi-Unit Intrafascicular
Recordings Using Neural Network Pattern-Recognition Techniques. IEEE Transactions on Biomedical
Engineering, VOL. 41, NO. 1(1994)
23. Simon Haykin, Neural Networks: A Comprehensive Foundation. Prentice Hall International, Inc(1998)
24. Chien-Cheng Lee, Pau-Choo Chung, Jea-Rong Tsai: Robust radial basis function neural networks. IEEE
Transactions on Systems, Man, and Cybernetics—part B: Cybernetics, vol. 29, no. 6(1999)
25. S . Chen, C. F. N. Cowan, P. M. Grant: Orthogonal least squares learning algorithm for radial basis function
networks. IEEE Transactions on Neural Networks, vol. 2, no. 2(1991)
26. J.Park, I. W. Sandberg: Universal approximation using radial-basis-function. Neural Computation 3, 246-257
(1991)
Authors
Authors did not want to publish their photos and bio-data.
47
International Journal of Advanced Science and Technology
Vol. 3, February, 2009
48