Related Experiment Video
Updated: May 25, 2025
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
Multi-feature fusion method combining brain functional connectivity and graph theory for schizophrenia classification
Chang Wang1, Yaning Ren1, Rui Zhang1
1Henan Collaborative Innovation Center of Prevention and treatment of mental disorder, The Second Affiliated Hospital of Xinxiang Medical University, Xinxiang, China; School of Medical Engineering, Xinxiang Medical University, Xinxiang, China; Henan Key Laboratory of Biological Psychiatry, Xinxiang, China; Engineering Technology Research Center of Neurosense and Control of Henan Province, Xinxiang, China; Xinxiang Engineering Technology Research Center of Intelligent Medical Imaging Diagnosis, Xinxiang, China.
Background:
The abnormalities in brain functional connectivity (FC) and graph topology (GT) in patients with schizophrenia (SZ) are unclear. Researchers proposed machine learning algorithms by combining FC or GT to identify SZ from healthy controls. The schizophrenia classification and neuroimaging markers screening using FC and GT feature fusion are blank.
Methods:
We proposed multi-feature fusion method combining functional connectivity and graph topology for schizophrenia classification and neuroimaging markers screening. Firstly, we acquired and preprocessed the private rs-fMRI data from the second affiliated hospital of Xinxiang Medical University in china. Secondly, we calculated the functional connectivity matrix and graph topology features. Thirdly, we used the two-sample t-test and the minimum absolute contraction selection operator (LASSO) to extract the features with statistical differences. Lastly, we used machine learning to classify schizophrenia and screen neuroimaging markers.
Results:
The result showed that the SVM model with the best feature (i.e., FC and GT) has the best performance (ACC = 0.935(95 percent confidence interval, 0.932 to 0.938), SEN = 0.920(95 percent confidence interval, 0.917 to 0.922), SPE = 0.950(95 percent confidence interval, 0.946 to 0.954), F1 = 0.935(95 percent confidence interval, 0.933 to 0.938), AUC = 0.935(95 percent confidence interval, 0.932 to 0.937)). We also found that the differences in FC and GT features are mainly located in the default network, the attention network, and the subcortical network. The feature strength of FC and GT showed a general decline in patients with SZ, and the node clustering coefficient of the thalamus and the FC of Putamen_L and Frontal_Mid_Orb_R showed an increase.
Conclusion:
It demonstrated that the multi-feature fusion has the advantage in distinguishing SZ from healthy individuals providing new insights into the underlying pathogenesis of SZ.
