Signal processing of the contingent negative variation in schizophrenia using multilayer perceptrons and predictive statistical diagnosis

An event related potential known as the contingent negative variation (CNV) was recorded from two sites from the brains of 20 medicated schizophrenics and 20 normal control subjects. The aim was to apply signal processing, artificial neural networks and statistical techniques to the CNV waveform to improve the understanding of schizophrenia and to develop a neurophysiological technique for its identification and monitoring. CNV recording sites were the vertex and from a point midline approximately 30mm anterior to the vertex (frontal). Three-layer multilayer perceptrons (MLPs) were used to discriminate between the CNV waveforms of the schizophrenics and normal subjects. Although the MLP technique was successful in discrimination, it did not provide a quantitative measure for the analysis. Furthermore, during the test phase it always classified the subjects into one of the two categories and did not provide an output for either type (unknown type). To improve the clinical diagnosis a discrimination technique based on predictive statistical diagnosis (PSD) was developed. The input parameters to the PSD were a time domain feature and three features obtained from the energy spectrum of the CNV waveform. The PSD output indicated the probability and the atypicality index of each subject belonging to one of the two groups. Discrimination accuracy of the PSD was 100% for normal subjects. Three schizophrenics could not be classified into either type, but the rest were identified correctly. T-tests carried out on the recorded CNV waveforms showed that the CNV waveform recorded from the vertex site in normal subjects is significantly different from that recorded from the frontal site; however this was not the case for schizophrenics.