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

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Major depressive disorder (MDD) diagnosis relies on clinical symptoms, lacking objective biomarkers.
• Brain connectome abnormalities are implicated in MDD pathophysiology, offering potential for diagnostic classification.
• Previous studies faced challenges due to small sample sizes, multi-site data inconsistencies, and complex graph structures.

Purpose of the Study:

• To develop and validate a novel graph convolution network with sparse pooling (GCNSP) for improved MDD classification using brain imaging data.
• To address challenges of multi-site data divergence and irregular connectome graph architectures.
• To identify hierarchical brain network dysfunction patterns associated with MDD.

Main Methods:

• A novel graph convolution network with sparse pooling (GCNSP) was proposed to learn hierarchical features from connectome graphs.
• The GCNSP model was applied to a large, multi-site functional MRI dataset (33 sites, 3335 subjects).
• Transfer learning was employed, pre-training GCNSP on a majority of sites and fine-tuning for cross-site classification.

Main Results:

• The GCNSP model achieved an average accuracy of 70.14% for MDD classification across 33 sites.
• Hierarchical dysfunction within the default mode network (DMN) was detected in MDD patients.
• Interactions between the DMN and frontoparietal network showed significant discriminative power between patients and controls.

Conclusions:

• The developed GCNSP offers an effective pipeline for multi-site diagnostic classification of MDD using neuroimaging.
• The study highlights the importance of hierarchical brain network dysfunction in MDD.
• Findings contribute to a better understanding of brain network alterations in neuropsychiatric disorders.