Spike Sorting Method Based on Two-Stage Radial Basis Function Networks - Advances in Computational Science and Engineering - page 103

Raw signals are superposed with white Gaussian noises in the measurement of dB,
which is the scalar of SNR that specifies the signal to noise ratio for better simulation of
real situation, using matlab function awgn(x, SNR). The scalar SNR specifies the
signal-to-noise ratio per sample in dB. This syntax assumes that the power of x is 0
dBW, and "dBW" representing a power level in decibels relative to 1 Watt.. Then we
can achieve the spike classification results in a range of different degrees of the envi-
ronmental noise.
3.2 Spike Detection
The typically used technique for detecting the action potentials is amplitude threshold
crossing, which extracts the action potentials when the amplitude exceeded a
threshold derived from the standard deviation of the signal from the baseline.
Methods such as matched filter are also widely used to work as a detector for action
potentials, which might achieve better performance than the amplitude threshold
crossing in the noisy environment [5]. Here, we use the method of matched filter as a
detector that maximized the SNR of the signal by a system whose transfer function is
the inversion of the typical action potential waveforms, which works better in our
experiments.
3.3 Spike Sorting
After spike detection, we get waveforms which are 90 points in length. According to
the waveforms obtained in experiment, we choose a part of 19 points as input vector
of RBF network (Fig. 3). To make train signals distribute uniformly in the data set,
first we use K-means to calculate centers of each class of spikes; then sort the vectors
according to its distance to the corresponding class center. From the data set, we get
one training vector, then one testing vector, then one training vector, etc. In this way
we get about 750 groups of training samples and 750 groups of testing samples for
each SNR. After training, we sort the spikes using the trained RBF network.
4 Experiment Results
The correct ratio of spike sorting under different SNR are shown in Table 1. Tab. 1
shows that when SNR is higher than -25 the correct ratio is higher than 95%. After
that, the correct ratio decrease fast and reaches about 70% when SNR = -40dB.
Table 2 shows a comparison between RBF network and BP network. From the table
we find the correct ratio of RBF network decrease earlier (when SNR = -30) than
BPNN (when SNR = -35). But RBF network achieve a higher correct ratio when SNR
reaches -35dB or even lower.