Bandpass Filter Based Artificial Neural Network for Neural Action Potential Detection

Abstract

Recognition of neural activity in a noisy environment is a critical point for scientists. The
bandpass filter is an essential and most straightforward method for neural spike classification;
however, this approach still very sensitive to the magnitude variances among the neural spike to be
beneficial in biomedical applications. Therefore, an enhanced bandpass filter using an artificial
neural network is suggested for improver the neural spike detection and sorting based on velocity
selective recording method. The proposed filter rejects the noise background and improves the
selectivity depend on fibre conduction velocity. Many approaches have been offered to implement
and upgrade the bandpass filter, such as finite impulse response and infinite impulse response
techniques. Adapting artificial neural network for digital filter designing approach is an advance
method could improve the filter characteristics due to its properties such as non-linearity and
flexibility. Different studies have been recommended for design this filter based on the artificial
network. However, random initial magnitudes are employing for training the neural network in these
approaches, as a result of this weakens, unstable performances and non-optimal design are produced.
Alternatively, a transfer function modification is applied in some studies for improving the digital
filter design, but this enhancement causes unwanted behaviour in most studies. This work proposes a
modern technique to solve the restrictions of past studies using an upgraded training approach. The
present method uses the coefficients of FIR bandpass filter as a target magnitude and the filter
characteristics such as filter order, cut-off frequencies, passband ripple, band stop ripple and
transition as an input value. A pre-existing (window) approach is employed to produce the
coefficients initial value and then reuse them to adjust the training for the artificial neural network
(error magnitude adjustment). The back propagation algorithm is applied to train the neural network
because it is a simple algorithm and sufficient error redaction. Minimizing the error magnitude
improve the filter performance such as the ability for reducing the passband ripple or boost the
stopband ripple individually, relying on the target FIR bandpass filter.