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

Qianqian Liao, Wuque Cai, Hongze Sun, Dongze Liu, Duo Chen, Dezhong Yao, Daqing Guo · 2025

Read A Brain-to-Population Graph Learning Framework for Diagnosing Brain Disorders by Qianqian Liao, Wuque Cai, Hongze Sun, Dongze Liu, Duo Chen, Dezhong Yao, Daqing Guo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-16096.txt.

Introduction

Qianqian Liao and co-authors present a 2025 paper on diagnosing brain disorders from scans. Graph methods compare how brain regions connect, yet they lean on fixed atlases and ignore rich detail inside them. Site and patient traits also blur the signal, which lowers trust in the results.

The two-stage framework first enriches each brain graph with atlas knowledge drawn from a large language model. An attention network then reshapes the graph before diagnosis begins. In the second stage, patient traits join the graph of the whole group to cut confounding effects. On autism, attention, and depression data sets, the method beat leading rivals and stayed easier to read. The authors pitch it for clinic use.

Catalog record

CREATORQianqian Liao, Wuque Cai, Hongze Sun, Dongze Liu, Duo Chen, Dezhong Yao, Daqing Guo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE70 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-16096

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Brain-to-Population Graph Learning Framework for Diagnosing Brain Disorders Qianqian Liao, Wuque Cai, Hongze Sun, Dongze Liu, Duo Chen, Dezhong Yao, Daqing Guo 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. 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 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. 1 Introduction Brain disorders—including neurodevelopmental conditions such as Autism Spectrum Disorder (ASD) ( Hodges et al. 2020 ) and Attention-Deficit/Hyperactivity Disorder (ADHD) ( Hinshaw 2018 ) , as well as psychiatric disorders like Major Depressive Disorder (MDD)—significantly impair cognitive, emotional, and social functioning. These conditions not only limit individuals’ potential but also place substantial burdens on families, communities, and healthcare systems ( Arora et al. 2018 ) . Early and accurate identification and intervention are critical to mitigate these impacts and improve outcomes for affected individuals ( Finlay-Jones et al. 2019 ) . Functional magnetic resonance imaging (fMRI) is a non-invasive technique for measuring brain activity through blood-oxygen-level-dependent (BOLD) signals ( Logothetis 2003 ) , offering critical insights into complex neurological disorders ( Park and Friston 2013 ; Arbabshirani et al. 2017 ; Tang et al. 2023 ) . As a widely used metric for quantifying synchronized BOLD activity, functional connectivity (FC) assesses interregional communication by analyzing temporal correlations between brain regions partitioned using predefined atlases ( Roland and Zilles 1994 ; Logothetis and Wandell 2004 ) . However, existing approaches to FC analysis fail to utilize the rich medical information in these atlases, such as functional characteristics and structural organization of brain regions, which could improve the interpretation of fMRI data ( Zhang et al. 2024 ) . Moreover, a major challenge in utilizing fMRI databases for neuroscience research lies in the presence of confounding effects, arising from site-related differences (e.g., MRI machine vendors) and population heterogeneity (e.g., age, gender) ( Gordon and Nelson 2021 ; Geerligs et al. 2017 ; Masi et al. 2017 ) . These confounds ( Rao et al. 2017 ; Brown et al. 2012 ) may cause models to rely on spurious correlations rather than clinically meaningful measures ( Ferrari et al. 2020 ) . Numerous computer-aided diagnosis methods have been developed based on FC analysis for diagnosing neurological disorders. Among these, graph neural networks (GNNs) have become a dominant approach owing to their powerful graph embedding capabilities. Building on this strength, recent works ( Scarselli et al. 2008 ) and ( Cui et al. 2022 ) have demonstrated the effectiveness of GNNs in analyzing both brain networks and population networks. However, two main limitations exist in current brain network GNNs: (1) the underutilization of essential information encoded in brain atlases, and (2) the sole focus on individual FC embeddings, neglecting the inter-individual relationships both within and across groups ( Wang et al. 2024b ) . Although the population GNNs incorporate non-imaging data to model relationships between subjects, they often rely on linking individuals from the same sites or with similar demographics ( Huang and Chung 2020 ; Zhang et al. 2023a ) . Given that GNNs operate through neighborhood aggregation ( Chen et al. 2024a ) , this approach amplifies the biases inherent in FC data, where confounding factors such as site and demographic variations already mask disease-relevant patterns. Moreover, despite the critical biological insights encoded in brain atlases, their potential remains underutilized due to the absence of methods for extracting interpretable and computationally efficient representations, coupled with the need for standardized and precise descriptions of brain regions. Recent advances in large language models (LLMs) have demonstrated promise in medical and neuroscience applications, offering potential solutions to bridge this gap ( Achiam et al. 2023 ) . These models can distill vast knowledge repositories into compact representations, serving as robust foundations for downstream tasks ( Bzdok et al. 2024 ) . Textual descriptions can additionally be encoded into high-dimensional vectors via embedding models, effectively capturing contextual relationships ( Jain and Chadha 2021 ) . Therefore, LLMs have been successfully deployed to generate precise explanations of medical concepts ( Yan et al. 2023 ) and compute semantic similarities between them ( Peng et al. 2024 ) . However, their direct application to information in the brain region remains largely unexplored. To address these aforementioned challenges, we propose a novel two-stage Brain-to-Population Graph Learning (B2P-GL) framework. In the first stage, termed brain representation learning, we leverage GPT-4 and the embedding model to construct a semantic-enhanced brain graph by calculating semantic similarities between brain regions. The graph is refined using the Adaptive Node Reassignment Graph Attention Network (ANR-GAT), which captures latent node relationships while improving the interpretability of the brain network. In the second stage, termed population disorder diagnosis, we construct a condition-based heterogeneous graph by connecting nodes with high similarity in brain features but different categories within one condition. This structure enables a Heterogeneous Graph Convolutional Network (HGCN) to aggregate cross-condition messages while mitigating confounding effects. We further integrate phenotypic data via gated fusion to boost diagnostic accuracy. The proposed framework is validated on three large-scale, publicly available brain disorder datasets (ABIDE I, ADHD-200, and Rest-meta-MDD), evaluating both the ANR-GAT model and the B2P-GL framework to ensure adaptability across broader applications. The main contributions of B2P-GL are as follows: 1. We propose the B2P-GL framework, which integrates a two-stage GNN to learn brain representations and population-level interactions, thereby improving both neuroscientific interpretability and diagnostic performance. 2. We introduce semantic brain region embeddings and adaptive node reassignment for richer brain graph representations, and a condition-based population graph leveraging phenotypic data to reduce confounding and guide HGCN toward disorder-relevant patterns. 3. We evaluate the proposed B2P-GL framework on three publicly available brain disorder datasets. Experimental results demonstrate that B2P-GL outperforms existing baselines, indicating its effectiveness and generalizability across diverse diagnostic tasks. 2 Related Work 2.1 GNN-based Neurological Disorders Diagnosis Computer-aided diagnosis using resting-state functional MRI (rs-fMRI) data has gained increasing interest, with approaches ranging from traditional machine learning methods like support vector machines ( Gallo et al. 2023 ) to deep learning techniques such as convolutional neural networks

Read the complete text