Source-linked AI summary
Interpretable Graph Neural Networks for Connectome-Based Brain Disorder Analysis
Hejie Cui, Wei Dai, Yanqiao Zhu, Xiaoxiao Li, Lifang He, Carl Yang
TL;DR
Connectome disorder analysis needs transparent graph-level explanations shared across subjects, while GNNs can overfit and existing explainers often produce subject-specific explanations. A connectome-oriented GNN backbone predicts disorders while a globally shared edge mask highlights disorder-specific biomarkers, and the mask-enhanced networks are used to further tune the backbone. About 9.7% relative improvement over the backbone is reported for IBGNN+, alongside good cross-dataset prediction and biomarkers coinciding with neuroscience findings.
Problem
Connectome disorder analysis needs transparent graph-level explanations shared across subjects, while GNNs can overfit and existing explainers often produce subject-specific explanations.
Source: §1 · p. 2Method
A connectome-oriented GNN backbone predicts disorders while a globally shared edge mask highlights disorder-specific biomarkers, and the mask-enhanced networks are used to further tune the backbone.
Source: §1 · p. 2, §2 · pp. 2, 4Results
Takeaways & Limitations
The work proposes a novel interpretable GNN framework for connectome-based brain disorder analysis.
Source: §5 · p. 8Small neuroimaging datasets may constrain the framework's effectiveness and deep-learning generalization ability.
Source: §5 · p. 9
Abstract
from arXiv · showhide
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.
1 Introduction
The paper introduces a connectome-oriented GNN framework that pairs disorder prediction with a globally shared explanation of disorder-specific biomarkers. Experiments across three brain-disorder datasets report strong prediction performance and biomarkers that coincide with neuroscience findings.
1 Introduction
- GNNs can capture complex connectome structures, but they may overfit and lack transparency; existing explainers often target node-level tasks or give each graph its own explanation.The paper notes that subjects with the same disorder share similar brain-network patterns.Source: §1 · pp. 1–2
- The framework combines a connectome-oriented prediction backbone with a globally shared mask that highlights disorder-specific ROIs and connections.The enhanced model, IBGNN+, jointly produces predictions and interpretations.Source: §1 · p. 2
- Across HIV, BP, and PPMI, the backbone performs well across neuroimaging modalities, generated biomarkers coincide with neuroscience findings, and combining the modules further boosts prediction.The same-disorder patterns motivate explanations shared across subjects.Source: §1 · p. 2
2 The Proposed Model
The proposed model combines a connectome-oriented GNN for disorder prediction with a globally shared explanation mask. It applies that mask to enhance and fine-tune the backbone, producing predictions and interpretations in a closed loop.
2 The Proposed Model
- The framework predicts each subject’s disorder and learns an interpretation matrix shared across subjects to highlight disorder-specific biomarkers.The matrix is applied to brain networks represented by ROIs and their weighted connections.Source: §2 · p. 2
- IBGNN handles brain-network edge weights by concatenating each node’s embedding and its neighbor’s embedding with their edge weight before message aggregation.This mechanism addresses the positive and negative correlation values used as brain-network edge weights.Source: §2 · pp. 2–3
- The explanation generator learns one edge mask across subjects by encouraging masked-graph predictions to agree with original predictions while regularizing explanation compactness and mask discreteness.The resulting mask can be used on testing graphs to investigate biomarkers and highlight disorder-specific salient connections.Source: §2 · p. 4
- Applying the learned mask to the original data highlights disorder-specific signals and allows the backbone to be fine-tuned; this enhancement is compatible with other backbones.The paper describes the combined prediction-and-interpretation pipeline as a way to further improve disorder prediction.Source: §2 · p. 4
3 Experiments
The framework was evaluated on three real-world brain-disorder datasets using different neuroimaging modalities. Both proposed models outperformed shallow and deep baselines, and the explanation-enhanced model further improved prediction performance.
Dataset acquisition and preprocessing
- The three datasets used different modalities and had groups balanced for age and gender, with the same image-acquisition procedure within each dataset.Results were averaged over ten-fold cross-validation.Source: §3 · pp. 5–6
- The HIV dataset used fMRI from 70 subjects and constructed networks with 90 cerebral regions; the BP dataset used DTI from 52 patients and 45 healthy controls.The HIV cohort comprised 35 early HIV patients and 35 seronegative controls; BP networks used 82 parcellated regions.Source: §3 · p. 5
- The PPMI DTI dataset included 754 subjects, 596 with Parkinson’s disease, and used 84 ROIs to reconstruct brain-network connectivity.The connectivity was reconstructed with deterministic second-order Runge-Kutta whole-brain tractography.Source: §3 · p. 5
Baselines
- IBGNN and IBGNN+ were compared with shallow graph methods and deep graph models, including brain-network-specific BrainNetCNN and BrainGNN.The shallow methods included M2E, MIC, MPCA, and MK-SVM; deep competitors included GAT, GCN, and PNA.Source: §3 · p. 6
Prediction performance
- IBGNN+ improved on its backbone IBGNN by about 9.7% relative improvement, while IBGNN achieved up to 11% absolute improvement over shallow models on BP.The reported overall results are in Table 1.Source: §3 · p. 6, Table 1 · p. 6
4 Interpretation Analysis
The interpretation analysis identifies salient ROIs and connections for HIV, bipolar disorder, and PPMI, and visualizes ROI importance and connection weights across groups and neural systems.
Salient ROIs
- ROIs were mapped to eight neural systems, and the salient ROI ranking used the sum of each node’s connected edge weights.The systems are VN, AN, BLN, DMN, SMN, SN, MN, and CCN.Source: §4 · pp. 6–7
Interpretation visualizations
- Fig. 2 colors ROIs by average importance within each group, with bright yellow indicating high and dark red low scores.Fig. 3 uses system colors for within-system edges, gray for cross-system edges, and edge width for explanation-graph weight.Source: Figure 2 · p. 7, Figure 3 · p. 8
Salient ROIs
- HIV salient ROIs were the anterior cingulate, paracingulate gyri, and inferior frontal gyrus; BP highlighted secondary visual cortex and medial-to-superior temporal gyrus; PPMI highlighted rostral and superior frontal gyri.The paper describes these observed ROIs as potential biomarkers for identifying disorders in each cohort.Source: §4 · p. 7
Important connections
- Patient explanations showed reduced or excluded DMN and VN interactions in HIV, sparser BLN connections in BP, and reduced SMN and DMN connectivity in PPMI.The paper relates these patterns to prior findings on the respective disorders.Source: §4 · p. 8, Figure 3 · p. 8
5 Conclusion
The paper presents an interpretable GNN framework combining a brain-network-oriented predictor with a globally shared explanation generator. Experiments report strong prediction performance and disorder-specific interpretations, while small neuroimaging datasets are identified as a limitation.
Conclusion
- The framework combines a brain-network-oriented GNN predictor with a globally shared explanation generator for connectome-based disorder analysis.The explanation generator highlights disorder-specific biomarkers.Source: §5 · p. 8
- Experiments on real-world neuroimaging datasets report superior prediction performance and validate disorder-specific interpretations from the generated mask.Source: §5 · p. 9
- Small neuroimaging datasets may constrain deep learning effectiveness and generalization; the paper proposes pre-training and transfer learning across datasets as future work.The proposed direction is intended to share information and explanations across cohorts.Source: §5 · p. 9