Auditory brainstem response classification using modular neural networks

A modular artificial neural-network system composed of two levels of modules has been developed for the classification of auditory brainstem response (ABR) to be used in electrophysiologic hearing-threshold determination. The first level consists of four neural-network modules, each of which is trained to recognize one of four basic ABR waveforms (high, mid, low and, no response). The final classifications (response, no response, and cannot determine) are decided by a second-level neural network or a rule-base system. The performance of the system was evaluated by computing Bayesian boundaries obtained by the nearest neighbor method. The results show that the modular system is trained faster than previous nonmodular networks and yields higher recognition rates. Eliminating stimulus and muscle artifacts by limiting input window further improves the recognition rate.