Machine Learning techniques for the diagnosis of Schizophrenia based on Event Related Potentials

Abstract

AbstractAntecedentThe diagnosis of schizophrenia could be enhanced with objective neurophysiological biomarkers, such as the event related potential features in conjunction with machine learning procedures. A previous work extracted features from event related responses to three oddball paradigms (auditory and visual P300, and mismatch negativity) for the discrimination of schizophrenic patients. They used several classifiers: Naïve Bayes, Support Vector Machine, Decision Tree, Adaboost and Random Forest. The best accuracy was obtained with Random Forest (84.7%).ObjectiveThe aim of this study was to examine the efficacy of Multiple Kernel Learning classifiers and Boruta feature selection method exploring different features for single-subject classification between schizophrenia patients and healthy controls.MethodsA cohort of 54 schizophrenic subjects and 54 healthy control subjects were studied. Three sets of features related to the event related potentials signal were calculated: Peak related features, Peak to Peak related features and Signal related features. The Boruta feature selection algorithm was used to evaluate its impact on classification accuracy. A Multiple Kernel Learning algorithm was applied to address schizophrenia detection.ResultsWe obtained a classification accuracy of 83% using Multiple Kernel Learning classifier with the whole dataset. This result in accuracy triangulates previous work and shows that the differences between schizophrenic patients and controls are robust even when different classifiers are used. Appling the Boruta feature selection algorithm a classification accuracy of 86% was yielded. The variables that contributed most to the classification were mainly related to the latency and amplitude of the auditory P300.ConclusionThis study showed that Multiple Kernel Learning can be useful in distinguishing between schizophrenic patients and controls. Moreover, the combination with the Boruta algorithm provides an improvement in classification accuracy and computational cost.