BrainIB: Interpretable Brain Network-Based Psychiatric Diagnosis With Graph Information Bottleneck.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Psychiatry

Background:

• Current psychiatric disorder diagnosis relies on subjective symptoms, lacking objective biological markers.
• Machine learning (ML) models using functional connectivity (FC) show promise but suffer from overfitting and poor generalization.
• Developing explainable and reliable brain biomarkers for clinical application remains a challenge.

Purpose of the Study:

• To develop a robust and generalizable machine learning framework for psychiatric disorder diagnosis.
• To identify informative brain network features for improved diagnostic accuracy.
• To discover reliable and clinically relevant brain biomarkers.

Main Methods:

• Proposed BrainIB, a graph neural network (GNN) framework utilizing the information bottleneck (IB) principle.
• Analyzed functional magnetic resonance imaging (fMRI) data to identify informative edges (subgraphs) in brain networks.
• Evaluated BrainIB against baseline and state-of-the-art methods on three psychiatric datasets.

Main Results:

• BrainIB achieved the highest diagnostic accuracy across three psychiatric datasets compared to existing methods.
• The framework demonstrated superior generalization to unseen data, mitigating overfitting issues.
• Identified subgraph biomarkers consistent with established clinical and neuroimaging findings.

Conclusions:

• BrainIB offers a promising approach for objective psychiatric disorder diagnosis using fMRI-based functional connectivity.
• The information bottleneck principle enhances model generalization and biomarker discovery.
• The framework has the potential for clinical application in psychiatric diagnostics.