Multi-Site Transfer Classification of Major Depressive Disorder: An fMRI Study in 3335 Subjects

Imaging-based automatic diagnosis of major depressive disorder (MDD) has received widespread attention in precision medicine. Increasing evidence suggests that the pathophysiology of MDD is associated with the abnormality in brain connectome, which could be an effective biomarker for classification. However, previous studies suffered from small number of samples and large multi-site imaging divergences, as well as irregular graph architectures of the connectome, which challenges the diagnostic classification of MDD. Here, we propose a novel graph convolution network with sparse pooling (GCNSP) to learn the hierarchical features of the connectome graph to improve MDD classification. We applied the model to a multi-site functional MRI sample (33 sites with 3335 subjects, the largest functional imaging dataset of MDD to date), and perform transfer learning classification for each site using the pre-trained GCNSP on remaining sites to fit cross-site divergences, achieving an average accuracy of 70.14%. Moreover, hierarchical dysfunction of default mode network (DMN) is detected by the GCNSP in the patients. The interaction between DMN and frontoparietal network exhibit high discriminative power between patients and controls. Accordingly, this study may provide an effective pipeline for multi-site diagnostic classification and improve our understanding of hierarchical clues of brain network dysfunction in neuropsychiatric disorders.