Random Subspace Ensemble Artificial Neural Networks for First-episode Schizophrenia Classification

Abstract

Computer-aided schizophrenia diagnosis is a difficult task that has been developing for last decades. Since traditional classifiers have not reached sufficient sensitivity and specificity, another possible way is combining the classifiers in ensembles. In this paper, we take advantage of random subspace ensemble method and combine it with multi-layer perceptron (MLP) and support vector machines (SVM). Our experiment employs voxel-based morphometry to extract the grey matter densities from 52 images of first-episode schizophrenia patients and 52 healthy controls. MLP and SVM are adapted on random feature vectors taken from predefined feature pool and the classification results are based on their voting. Random feature ensemble method improved prediction of schizophrenia when short input feature vector (100 features) was used, however the performance was comparable with single classifiers based on bigger input feature vector (1000 and 10000 features).