Classification of drug-induced behaviors using a multi-layer feed-forward neural network

Measurement of laboratory animal motor behavior is an important part of many studies of experimental manipulations of the central nervous system. Current automated data collection and analysis systems are limited in the number of behaviors that can be monitored and quantified simultaneously. This paper describes a signal analysis technique that when used to analyze the data from a modified Stoelting electronic activity monitor is capable of classifying multiple behavior categories automatically. In this technique, the output signal from the motility monitor is fixed-length segmented and feature extraction is performed, calculating the Fourier transform and power spectrum of each data segment. An error back-propagation neural network, implemented on a microcomputer, is used to perform behavior classification of the segment power spectra. The technique provides a high degree of accuracy in automatic behavior classification and should prove useful in the quantitative assessment of behavior.