Brain Network Analysis Of Schizophrenia Patients Based On Hypergraph Signal Processing

Schizophrenia is a serious neuropsychiatric disease,currently mainly diagnosed by experts based on clinical symptoms.Researches on medical imaging and graph theory have found that the topological properties of patients’ brain functional networks can significantly change.Hypergraph-based functional networks can describe the interactions between multiple brain regions,and this high order relationship is crucial for exploring the pathogenesis of neurological diseases.This work constructs a brain network hypergraph spectrum analysis framework based on hypergraph signal processing,and uses hypergraph learning to analyze the differences of topological attributes between groups.The main contents are as follows,(1)A brain network spectrum analysis framework based on hypergraph signal processing.This work proposes a new hyperedge weight calculation method,which combines sparse representation and Pearson correlation to construct a weighted brain hypernetwork described by adjacency tensors.On the basis of tensor decomposition,this work uses hypergraph signal processing to analyze and compare the frequency spectrum of schizophrenia patients and healthy control groups.It is found that there are more high-frequency components in the frequency spectrum of patients than in the control group,and the average amplitude is significantly greater than in the control group.This work innovatively uses hypergraph spectrum as classification features,achieving high classification accuracy on two public datasets,proving the effectiveness of the proposed method.(2)Topological metric difference analysis based on group-level hypergraph learning.Most of the hypergraphs constructed in current work are based on individuals and cannot analyze changes caused by diseases at the group level.To overcome this limitation,this work uses hypergraph learning to identify brain regions related to diseases.We construct a graph based functional network for each subject,derive graph metrics and construct a group level hypergraph based on the similarity of graph metrics.Through hypergraph learning,we obtaine the weights of each hyperedge.We evaluate the framework on datasets of two different diseases and test the versatility of the method.Experiments have found that hyperedges with larger and smaller weights have good discrimination,and using the corresponding features of these hyperedges can achieve good classification results.