BPI-GNN: Interpretable brain network-based psychiatric diagnosis and subtyping.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Psychiatric disorders like major depressive disorder (MDD) and autism spectrum disorder (ASD) are heterogeneous, complicating diagnosis and treatment.
• Clinical heterogeneity impedes progress in precision psychiatry and effective treatment strategies.

Purpose of the Study:

• To develop an interpretable graph neural network (GNN) framework, BPI-GNN, for analyzing functional magnetic resonance imaging (fMRI) data.
• To enhance the discrimination of psychiatric patients from healthy controls and identify biologically meaningful subtypes of psychiatric disorders.

Main Methods:

• Proposed BPI-GNN, an interpretable GNN framework leveraging prototype learning for fMRI analysis.
• Introduced a novel prototype subgraph generation process and used total correlation (TC) for pattern independence.
• Evaluated BPI-GNN against 11 classification methods on three psychiatric datasets.

Main Results:

• BPI-GNN achieved the highest diagnostic accuracy across all tested datasets, outperforming 11 popular brain network classification methods.
• Identified brain-based subtypes of psychiatric disorders with significant clinical relevance, correlating with symptom and gene expression profiles.
• Discovered subtype biomarkers consistent with current neuroscientific understanding.

Conclusions:

• BPI-GNN effectively discriminates psychiatric patients and identifies clinically relevant subtypes, offering a promising tool for precision psychiatry.
• The framework's interpretability and high accuracy advance the understanding and classification of complex psychiatric conditions.
• Identified subtypes and biomarkers provide a foundation for developing targeted diagnostic and therapeutic strategies.