Potencial diagnóstico não invasivo da esquizofrenia com redes neurais artificiais: EEG e estímulos visuais acromáticos

Abstract

Schizophrenia is a mental illness that a effcts about 1% of the world's population. The diagnosis of this illness is still dependent of direct interview with patients. One of the main challenges is to create an effective non-invasive methodology to classify schizophrenia. This study aimed at developing a classification test of schizophrenia through Artificial Neural Network (ANN) based on the EEG (Visual Provoked Potential) response over visual stimulation. Thirty-six volunteers participated in the study, 18 schizophrenics and 18 non-schizophrenics. We used an EEG system (actiCHamp) with 32 active electrodes during a checkerboard stimuli. Our ANN was a Multilayer Perceptron and its architecture was composed of an input layer with ten features, a hidden layer with three neurons and an output layer with two neurons representing the control and schizophrenics classes. Furthermore, cross validation (K = 2) was used as validation method. The results show an accuracy of 83.333% for the general model. The following results were observed for the Schizophrenics: True-Positive-Rate = .889; False-Positive = .222; Precision = .800; Recall = .889. For Non-schizophrenic the next values were obtained: True-Positive-Rate = .778; False-Positive = .111; Precision = .875; Recall = .778. A further experiment was performed by dividing the instainces into a 66% training set and a 33% test set, where we obtained a mean accuracy of 91.7%. We believe that the used stimulus increases quality and accuracy; while decreases the time of data collection. These features are important aspects for research procedures with clinical populations. Studies like that may contribute to future diagnosis of schizophrenia, a heterogeneous and chronic disease whose diagnosis is still inaccurate.