Biomarkers Selection of Abnormal Functional Connections in Schizophrenia with ℓ_(2,1-2)-Norm Based Sparse Regularization Feature Selection Method

摘要

Schizophrenia (SZ) is a serious mental disease involving multiple symptoms such as sensation, cognition, emotion and memory. Recent studies have shown that schizophrenia is related to abnormal functional connections in brain regions. Functional magnetic resonance imaging (fMRI) has become a powerful tool to examine the abnormal connectivity of brain networks in SZ, which provides an approach to bring psychiatry from subjective description into objective and tangible brain-based measures. However, there are many irrelevant and redundant features in fMRI. How to eliminate redundant and irrelevant features and locate accurately the biomarkers for schizophrenia from fMRI data are very important for diagnosis and further study of SZ. In this paper, the ℓ_(2,1-2)-norm, which is a kind of nonconvex and Lipschitz continuous norm, is introduced as a regularization term to the embedded feature selection method. The ℓ_(2,1-2)-norm sparse regularization feature selection method can be used to select features that are crucial for discrimination, and then explore the abnormal functional connections of schizophrenia with fMRI data. Our results showed that the abnormal functional connections are related to superior parietal gyrus, parahippocampal gyrus, caudate nucleus and middle occipital gyrus. The changes of functional connections involved in these brain regions may result to incorrect information processing in the brain, which provide further evidence for the cognitive disorder of schizophrenia.