Comprehensive Summary
This systematic review evaluates how graph neural networks (GNNs) have been used with resting-state fMRI data to identify potential biomarkers for major psychiatric disorders—including ADHD, ASD, MDD, and schizophrenia. After screening 65 studies, the authors find that while many GNN models achieve high classification accuracy, the specific biomarkers they report are highly inconsistent across studies, even when analyzing the same disorder and using similar methods. The review highlights that current approaches to evaluating biomarker robustness rely heavily on subjective comparisons with prior literature rather than objective metrics. To address this, the paper proposes a unified prediction–attribution–evaluation framework and emphasizes the need for consistent standards, better graph construction practices, and more rigorous assessment of attribution methods.
Outcomes and Implications
The review shows that GNNs and modern model-explainability tools (such as feature attribution methods) have significant potential to advance biomarker discovery in psychiatry, but their clinical reliability remains limited. Because GNNs can directly model complex functional connectivity patterns, they offer a more biologically meaningful way to detect subtle neural signatures of psychiatric disorders. However, the wide variability in biomarkers across studies suggests that current attribution methods may be sensitive to model design choices, dataset differences, or noise inherent in fMRI. As a result, the paper stresses that future technological development must prioritize robustness, reproducibility, and objective evaluation metrics. Improved explainability techniques, better harmonization of fMRI data, and more biologically grounded GNN architectures could ultimately make computational biomarkers more trustworthy—paving the way toward clinical adoption and more personalized, brain-based psychiatric diagnostics.