ARTIFICIAL NEURAL-NETWORK AND SPECTRUM ANALYSIS-METHODS FOR DETECTINGBRAIN DISEASES FROM THE CNV RESPONSE IN THE ELECTROENCEPHALOGRAM

Two methods of identifying schizophrenia, Parkinson's disease (PD), an d Huntington's disease (HD) are described. The methods are based on th e analysis of the contingent negative variation (CNV), an event relate d potential (ERP) in the electroencephalogram. The first method involv es spectrum analysis of the CNV and discriminant analysis of the Fouri er harmonic frequency components. The other method involves the applic ation of supervised learning artificial neural networks to the CNV fea tures obtained in the time domain. Additionally, unsupervised artifici al neural networks were used to presymptomatically assess the risk of HD. Sensitivities and specificities lie between 0.81 and 1.0 with low false positive rates (0 to 0.13) for differentiating between disease a nd normal data, and between disease data, dependent on disease and met hod. The preferred method for disease differentiation for accuracy and ease of application is the multilayer perceptron Using Kohonen and AR T networks for detecting abnormal CNVs in subjects at risk of HD (ARs) eight abnormals are identified in agreement with the prediction of ri sk derived from a published risk table. In addition, one of the abnorm als has since developed symptomatic Huntington's disease. The recommen ded method is to combine the results of the Kohonen method with an ART 2 and a modified ART1 network.