A brain-to-population graph learning framework for diagnosing brain disorders.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Current graph-based brain disorder diagnosis methods rely heavily on predefined atlases.
• These methods often ignore rich atlas information and confounding factors like site and phenotype variability.
• This limits diagnostic accuracy and personalization.

Purpose of the Study:

• To propose a two-stage Brain-to-Population Graph Learning (B2P-GL) framework.
• To integrate semantic similarity of brain regions and condition-based population graph modeling.
• To enhance brain disorder diagnosis by addressing limitations of existing methods.

Main Methods:

• Stage 1: Brain representation learning using GPT-4 for atlas knowledge integration and adaptive node reassignment graph attention network for brain graph refinement.
• Stage 2: Population disorder diagnosis incorporating phenotypic data for population graph construction and feature fusion.
• Utilized ABIDE I, ADHD-200, and Rest-meta-MDD datasets for validation.

Main Results:

• The B2P-GL framework demonstrated superior prediction accuracy compared to state-of-the-art methods.
• The framework enhanced the interpretability of diagnostic models.
• Effectively mitigated confounding effects from site and phenotype variability.

Conclusions:

• B2P-GL offers a reliable and personalized approach to brain disorder diagnosis.
• The framework advances the clinical applicability of AI in neuroscience.
• Successfully integrated advanced AI techniques with neuroimaging data for improved diagnostic outcomes.