A Brain-to-Population Graph Learning Framework for Diagnosing Brain Disorders

TL;DR

The paper tackles the challenge of diagnosing brain disorders from rs-fMRI by exploiting atlas-informed semantic knowledge and addressing site and demographic confounds. It introduces a two-stage Brain-to-Population Graph Learning (B2P-GL) framework that first builds a semantic-enhanced brain graph via GPT-4-derived region embeddings refined with an Adaptive Node Reassignment GAT, then constructs a condition-based population graph processed by a Heterogeneous Graph Convolution with gated phenotypic fusion. Key contributions include semantic brain region embeddings, adaptive node reassignment for richer brain representations, a cross-condition population graph to mitigate confounds, and a gated fusion mechanism to integrate phenotypic data, validated on ABIDE I, ADHD-200, and Rest-meta-MDD with superior accuracy and interpretability. The approach yields robust, personalized diagnostic patterns and demonstrates clinical relevance by revealing disorder-specific and shared neurofunctional signatures, with potential to improve clinical applicability across diverse brain disorders.

Abstract

Recent developed graph-based methods for diagnosing brain disorders using functional connectivity highly rely on predefined brain atlases, but overlook the rich information embedded within atlases and the confounding effects of site and phenotype variability. To address these challenges, we propose a two-stage Brain-to-Population Graph Learning (B2P-GL) framework that integrates the semantic similarity of brain regions and condition-based population graph modeling. In the first stage, termed brain representation learning, we leverage brain atlas knowledge from GPT-4 to enrich the graph representation and refine the brain graph through an adaptive node reassignment graph attention network. In the second stage, termed population disorder diagnosis, phenotypic data is incorporated into population graph construction and feature fusion to mitigate confounding effects and enhance diagnosis performance. Experiments on the ABIDE I, ADHD-200, and Rest-meta-MDD datasets show that B2P-GL outperforms state-of-the-art methods in prediction accuracy while enhancing interpretability. Overall, our proposed framework offers a reliable and personalized approach to brain disorder diagnosis, advancing clinical applicability.

Figures (9)

• Figure 1: The framework of our proposed Brain-to-Population Graph Learning Network. The upper section depicts the first stage, which constructs semantic-enhanced brain graphs and learns brain representations using the Adaptive Node Reassignment Graph Attention Network (ANR-GAT). The lower section illustrates the second stage, building a condition-based population graph and capturing meaningful patterns via a Heterogeneous Graph Convolutional Network (HGCN) and a gated fusion mechanism to integrate phenotypic data.
• Figure 2: Illustration of semantic similarity generation for brain regions. Functional and structural descriptions (e.g., for the left amygdala) are encoded using a text encoder to produce embeddings. These embeddings are then used to compute semantic similarity between regions.
• Figure 3: Prompt for generating neuroanatomical descriptions.
• Figure 4: Prompt for structured extraction of neuroanatomical and functional concepts from text.
• Figure 5: Visualization of brain feature embeddings using t-SNE. Compared with stage 1 embeddings derived from individual graphs and stage 2 embeddings updated through HGCN without phenotypic fusion.