Multimodal Imaging Genetic Analysis of Schizophrenia Based on Sparse Representation and Dictionary Learning

Schizophrenia is a highly hereditary illness, and there is no gold standard for the diagnosis of this disease. The clinical diagnosis of this disease lacks objective biomarkers. In order to extract biomarkers, machine learning methods are used. Most of machine learning methods only consider single-modality imaging data or genetic data, which makes the extracted biomarkers inaccurate. In this paper, we propose a new group and orthogonal constraints joint sparse representation and dictionary learning (GOJSDL) model to improve the accuracy and diversity of the schizophrenia biomarkers. Specifically, the proposed model maps multimodal imaging genetic data of schizophrenia patients to a common dictionary matrix to take full advantage of the complementary information in multimodal data. Then, to reduce the collinearity of atoms in the common dictionary matrix preventing ambiguity in feature selection, we incorporate orthogonal constraints into the SDL model. In addition, we introduce group structure as prior constraint into the SDL model to realize the sparseness of the features at the group level. We verify the effectiveness and state-of-the-art of the proposed model based on simulated datasets. Subsequently, we apply the GOJSDL model to a real image genetic dataset of schizophrenia, and identify abnormal brain regions (e.g., Hippocampus and Fusiform_R) from the functional magnetic resonance imaging (fMRI) dataset, disease-related risk genes (e.g., NFX1, CDH4) from the single nucleotide polymorphism (SNP) data and disease-related environmental factors (such as ADNP, RAET1L) from the DNA methylation data. Through literature review and functional analysis, it is shown that these biomarkers are closely related to schizophrenia. The discovery of these biomarkers helps us to further explore the pathogenesis of schizophrenia.