[Paper Review] Interpretable Graph Neural Networks for Connectome-Based Brain Disorder Analysis

Hejie Cui, Wei Dai+4 more|arXiv (Cornell University)|Jun 30, 2022

Functional Brain Connectivity Studies4 citations

TL;DR

This paper proposes an interpretable graph neural network framework, IBGNN, for connectome-based brain disorder analysis that jointly predicts brain disorders and identifies disorder-specific biomarkers. By combining a brain-network-aware GNN backbone with a globally shared explanation generator, the model enhances prediction performance and reveals biologically meaningful Regions of Interest and connections across HIV, bipolar disorder, and Parkinson’s disease datasets.

ABSTRACT

Human brains lie at the core of complex neurobiological systems, where the neurons, circuits, and subsystems interact in enigmatic ways. Understanding the structural and functional mechanisms of the brain has long been an intriguing pursuit for neuroscience research and clinical disorder therapy. Mapping the connections of the human brain as a network is one of the most pervasive paradigms in neuroscience. Graph Neural Networks (GNNs) have recently emerged as a potential method for modeling complex network data. Deep models, on the other hand, have low interpretability, which prevents their usage in decision-critical contexts like healthcare. To bridge this gap, we propose an interpretable framework to analyze disorder-specific Regions of Interest (ROIs) and prominent connections. The proposed framework consists of two modules: a brain-network-oriented backbone model for disease prediction and a globally shared explanation generator that highlights disorder-specific biomarkers including salient ROIs and important connections. We conduct experiments on three real-world datasets of brain disorders. The results verify that our framework can obtain outstanding performance and also identify meaningful biomarkers. All code for this work is available at https://github.com/HennyJie/IBGNN.git.

Motivation & Objective

• To address the lack of interpretability in deep learning models for brain disorder diagnosis.
• To identify common, disorder-specific brain network patterns across subjects, rather than individual-specific explanations.
• To improve disease prediction performance using a unified, globally shared explanation mask.
• To validate the biological plausibility of identified biomarkers against existing neuroscience literature.
• To develop a framework that integrates subject-specific connectomes with group-level interpretations for multi-level analysis.

Proposed method

• The backbone model, IBGNN, uses an edge-weight-aware message passing mechanism that incorporates both node features and weighted edge connections between brain regions.
• Node representations are updated via a message aggregation rule that sums messages from neighbors, including self-connections, using a non-linear activation function.
• A globally shared explanation mask is learned across all subjects to highlight disorder-specific Regions of Interest (ROIs) and important connections.
• The explanation mask is applied to enhance the original brain network, creating a masked graph that is fine-tuned with the backbone model to form IBGNN+.
• The explanation generator produces a single, shared interpretation matrix that identifies salient ROIs and connections common across patients with the same disorder.
• The framework is trained end-to-end, enabling simultaneous prediction and interpretation, with performance evaluated on three real-world neuroimaging datasets.

Experimental results

Research questions

• RQ1 Can a GNN-based model achieve high performance in connectome-based brain disorder prediction while remaining interpretable?
• RQ2 Do the globally shared explanations identify biologically plausible biomarkers that align with existing neuroscience findings?
• RQ3 Can the integration of explanation masks improve the predictive performance of the GNN backbone?
• RQ4 Are the identified salient ROIs and connections consistent across different brain disorders such as HIV, bipolar disorder, and Parkinson’s disease?
• RQ5 Can the model reveal group-level network patterns that are robust to individual variability in image quality?

Key findings

• The IBGNN backbone model achieved strong performance across three real-world datasets (HIV, BP, PPMI), demonstrating generalization across different neuroimaging modalities.
• The globally shared explanation mask successfully identified disorder-specific biomarkers, such as the inferior frontal gyrus in HIV patients and the rostral middle frontal gyrus in Parkinson’s disease patients.
• The explanation subgraph revealed significant disruptions in the default mode network (DMN) and visual network (VN) in HIV patients, consistent with known visual processing deficits.
• For bipolar disorder, the model highlighted reduced connectivity within the dorsal attention network (BLN), aligning with clinical findings of parietal lobe dysfunction.
• In Parkinson’s disease, decreased connectivity in the sensorimotor network (SMN) and DMN was observed, confirming prior neuroimaging studies on motor and cognitive decline.
• The IBGNN+ model, which fine-tunes the backbone using the explanation-enhanced graph, showed further performance gains over the base model, validating the utility of the explanation module.