Discovering robust biomarkers of psychiatric disorders from resting-state functional MRI via graph neural networks: A systematic review

TL;DR

This systematic review examines how graph neural networks (GNNs) applied to resting-state fMRI (rs-fMRI) have been used to identify potential biomarkers for psychiatric disorders. It synthesizes 65 studies across ADHD, ASD, MDD, SZ, and transdiagnostic contexts, highlighting that although model performance is often high, attribution robustness and cross-study reproducibility of biomarkers remain limited. The authors propose a comprehensive taxonomy of GNN architectures tailored to fMRI data, a taxonomy of attribution methods ( self-interpretable and post-hoc ), and a framework for evaluating attribution robustness (prediction–attribution–evaluation). They find region-level biomarkers show somewhat more reproducibility than edge-level features, but stresses the need for standardized reporting, objective evaluation metrics, and benchmarking against state-of-the-art GNNs. The work outlines a path forward toward robust, generalisable brain biomarkers via GNNs, including data harmonisation, multi-modal integration, temporal GNNs, and regression or transdiagnostic approaches beyond binary disease classification.

Abstract

Graph neural networks (GNN) have emerged as a popular tool for modelling functional magnetic resonance imaging (fMRI) datasets. Many recent studies have reported significant improvements in disorder classification performance via more sophisticated GNN designs and highlighted salient features that could be potential biomarkers of the disorder. However, existing methods of evaluating their robustness are often limited to cross-referencing with existing literature, which is a subjective and inconsistent process. In this review, we provide an overview of how GNN and model explainability techniques (specifically, feature attributors) have been applied to fMRI datasets for disorder prediction tasks, with an emphasis on evaluating the robustness of potential biomarkers produced for psychiatric disorders. Then, 65 studies using GNNs that reported potential fMRI biomarkers for psychiatric disorders (attention-deficit hyperactivity disorder, autism spectrum disorder, major depressive disorder, schizophrenia) published before 9 October 2024 were identified from 2 online databases (Scopus, PubMed). We found that while most studies have performant models, salient features highlighted in these studies (as determined by feature attribution scores) vary greatly across studies on the same disorder. Reproducibility of biomarkers is only limited to a small subset at the level of regions and few transdiagnostic biomarkers were identified. To address these issues, we suggest establishing new standards that are based on objective evaluation metrics to determine the robustness of these potential biomarkers. We further highlight gaps in the existing literature and put together a prediction-attribution-evaluation framework that could set the foundations for future research on discovering robust biomarkers of psychiatric disorders via GNNs.

Figures (9)

• Figure 1: Flowchart detailing the selection process of this review.
• Figure 2: Summary of the key components of typical GNNs used on fMRI datasets.
• Figure 3: Our proposed taxonomy of state-of-the-art GNN models customised for fMRI datasets.
• Figure 4: Taxonomy of attributors applicable to GNNs. Generally, post-hoc methods (typically 'black box') have a larger range of algorithms than self-interpretable (typically 'grey box') ones. Methods that have not been explored in fMRI studies are highlighted in yellow.
• Figure 5: A subset of the Co-12 properties nauta2023anecdotal used to evaluate feature attribution scores produced by attributors, deemed to be relevant for biomarker discovery from fMRI datasets.