Classification of schizophrenia patients on lattice computing resting-state fMRI features

Abstract

Currently there are many efforts to find neurological biomarkers that can be extracted from resting state fMRI data. In this paper we concentrate on a study about the discrimination between schizophrenia patients and healthy control, as well as the discrimination of subpopulations of schizophrenia patients with and without auditory hallucinations. Specifically, we compute scalar measures of the high dimensional fMRI voxel time series, carrying out feature selection, feature extraction and classification by SVM over them. The dimensionality reduction is formalized as a supervised h-function proposed in recent Multivariate Mathematical Morphology approaches. It is computed as the recall error of a Lattice Auto-Associative Memory. Using as background and foreground seeds the average time series of regions of interest extracted from the brain ventricles and auditory cortex, respectively. Results on a database of healthy controls and schizophrenia patients with and without auditory hallucinations show that the approach can provide accurate discrimination between these populations.