Characterizing Psychiatric Disorders Through Graph Neural Networks: A Functional Connectivity Analysis of Depression...

Area of Science:

• Neuroscience
• Psychiatry
• Artificial Intelligence in Medicine

Background:

• Major depressive disorder (MDD) and schizophrenia (SZ) are severe psychiatric disorders with widespread brain network disruptions.
• Understanding the commonalities and differences in large-scale network alterations is crucial for elucidating shared and disorder-specific neural mechanisms.
• Advanced deep learning, particularly graph neural networks (GNNs), offers novel methods for analyzing complex brain connectivity.

Purpose of the Study:

• To employ deep learning techniques, specifically GNNs, to identify common and distinct large-scale brain network patterns in MDD and SZ.
• To gain insights into the shared and disorder-specific neural underpinnings of these psychiatric conditions.
• To explore the potential of GNNs for diagnostic classification and biomarker discovery in psychiatric disorders.

Main Methods:

• Application of state-of-the-art GNN architectures, including the attention-based hierarchical pooling GNN (SAGPool), to a multisite resting-state fMRI dataset.
• Classification of major depressive disorder (MDD) and schizophrenia (SZ) patients against healthy controls (HCs).
• Utilized a perturbation-based explainability method to identify key functional connections influencing model predictions.

Main Results:

• The SAGPool GNN model achieved high classification accuracies: 71.50% for MDD and 75.65% for SZ.
• Distinct patterns of large-scale network disruption were identified: default mode network (DMN) alterations in MDD and ventral attention network (VAN) alterations in SZ.
• Specific identified functional connections correlated significantly with clinical symptoms in SZ patients, including positive and general symptoms (PANSS).

Conclusions:

• Graph neural networks are effective tools for uncovering complex brain connectivity patterns in psychiatric disorders.
• The study provides novel insights into the distinct neural mechanisms underlying MDD and SZ.
• These findings highlight the potential of graph-based models for diagnostic classification and biomarker discovery in psychiatric research.