Extraction of large-scale structural covariance networks from grey matter volume for Parkinson’s disease classification

Objectives: To identify disease-related spatial covariance patterns of grey matter volume as an aid in the classification of Parkinson’s disease (PD). Methods: Seventy structural covariance networks (SCNs) based on grey matter volume covariance patterns were defined using independent component analysis with T1-weighted structural MRI scans (discovery sample, 70 PD patients and 70 healthy controls). An image-based classifier was constructed from SCNs using a multiple logistic regression analysis with a leave-one-out cross-validation-based feature selection scheme. A validation sample (26 PD patients and 26 healthy controls) was further collected to evaluate the generalization ability of the constructed classifier. Results: In the discovery sample, 13 SCNs, including the cerebellum, anterior temporal poles, parahippocampal gyrus, parietal operculum, occipital lobes, supramarginal gyri, superior parietal lobes, paracingulate gyri and precentral gyri, had higher classification performance for PD. In the validation sample, the classifier had moderate generalization ability, with a mean sensitivity of 81%, specificity of 69% and overall accuracy of 75%. Furthermore, certain individual SCNs were also associated with disease severity. Conclusions: Although not applicable for routine care at present, our results provide empirical evidence that disease-specific, large-scale structural networks can provide a foundation for the further improvement of diagnostic MRI in movement disorders. Key Points: • Disease-specific, large-scale SCNs can be identified from structural MRI. • A new network-based framework for PD classification is proposed. • An SCN-based classifier had moderate generalization ability in PD classification. • The selected SCNs provide valuable functional information regarding PD patients.