A neural network design for event-related potential diagnosis
Fred Y. Wu, Jeremy D. Slater, Lawrence S. Honig, Richard Eugene Ramsay · Computers in Biology and Medicine · 1993
Many diseases resulting in neuropsychological impairment show abnormalities by EEG (electroencephalogram) tests. However, EEG analysis is complicated by a wide spectrum of normal patterns. Cerebral evoked potentials are stimulus-induced, averaged EEG potentials that have been found useful in patients with dementing illness. A recent report using the P300 auditory evoked potential shows that simple latency and waveform criteria result in classification accuracy of 65% for multiple sclerosis (MS) patients versus 91% for control subjects. Analysis on the same data set was performed using an artificial neural network (ANN) and a nearest neighbor (NN) classifier. An ANN classifier demonstrated a classification accuracy of 75% versus 87% on MS and control subject groups. Thus prediction accuracy was improved on average, compared with that obtained by NN classifiers or P300 statistical analysis. The classification strategy discovered by a trained ANN was analyzed by a weight pattern analysis method and compared with the P300 latency criteria.