Research map / Graph learning / Graph self-supervised learning
BrainOOD: Out-of-distribution Generalizable Brain Network Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BrainOOD, a GIB-based GNN with feature masking, high-pass reconstruction, discrete edge sampling and an alignment loss for OOD-generalizable and interpretable brain network classification, plus the first site-holdout OOD brain network benchmark on ABIDE and ADNI.
Problem
Brain network classifiers degrade under multi-site distribution shifts when applied to unseen sites, and existing graph OOD/causal-subgraph methods ignore node-feature noise and fail to give consistent, interpretable ROI selections for neurological disorders.
Approach
Extends the graph information bottleneck with a learnable node-feature mask trained by reconstruction through a high-pass GNN, discrete Gumbel-based edge sampling for subgraph extraction, and an alignment loss that encourages the same connections to be selected across all brain networks in a batch.
Key results
- BrainOOD achieves 64.81% OOD accuracy on ABIDE versus 60.38% for the best baseline (BrainNetCNN), a 7.34% relative gain as reported in the text.
- On ADNI 6-class, BrainOOD reaches 64.07% OOD accuracy and 64.80% overall accuracy, the best among compared methods.
- Ablation on ABIDE shows removing the entropy, reconstruction, or alignment loss, or using unmasked features or the original adjacency, lowers overall accuracy from 63.95% to between 60.10% and 63.12%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply BrainOOD with the GIN backbone rather than the strongest listed baseline | ABIDE OOD classification · ABIDE OOD | accuracy: 60.38 ± 7.07 -> 64.81 ± 9.01 | Helped |
| Apply BrainOOD with the GIN backbone rather than the strongest listed baseline | ADNI 5-class OOD classification · ADNI 5-class OOD | accuracy: 60.93 ± 13.27 -> 62.26 ± 15.83 | Helped |
| Apply BrainOOD with the GIN backbone rather than the strongest listed baseline | ADNI 6-class OOD classification · ADNI 6-class OOD | accuracy: 60.79 ± 13.97 -> 64.07 ± 14.99 | Helped |
| Add entropy loss (ablation) | BrainOOD on ABIDE · ABIDE | OOD accuracy: 63.70 ± 4.53 -> 64.81 ± 9.01 | Helped |
| Add reconstruction loss (ablation) | BrainOOD on ABIDE · ABIDE | OOD accuracy: 61.37 ± 7.13 -> 64.81 ± 9.01 | Helped |
| Add alignment loss (ablation) | BrainOOD on ABIDE · ABIDE | OOD accuracy: 60.43 ± 5.45 -> 64.81 ± 9.01 | Helped |
| Use masked features X' rather than unmasked features X (ablation) | BrainOOD on ABIDE · ABIDE | OOD accuracy: 62.26 ± 5.68 -> 64.81 ± 9.01 | Helped |
| Use sampled adjacency A' rather than original adjacency A (ablation) | BrainOOD on ABIDE · ABIDE | OOD accuracy: 55.40 ± 8.95 -> 64.81 ± 9.01 | Helped |
| Use the selected hyperparameters k=5, λ1=0.01, λ2=0.1, λ3=0.5 instead of k=1 (ablation) | BrainOOD on ABIDE · ABIDE | overall accuracy: 61.95 ± 4.54 -> 63.95 ± 4.65 | Helped |
| Increase sampling count from k=5 to k=10 (ablation) | BrainOOD on ABIDE · ABIDE | overall accuracy: 63.95 ± 4.65 -> 62.90 ± 4.67 | Hurt |
| Increase sampling count from k=5 to k=20 (ablation) | BrainOOD on ABIDE · ABIDE | overall accuracy: 63.95 ± 4.65 -> 61.59 ± 3.57 | Hurt |
| Set λ2 to 1.0 rather than the selected 0.1 (ablation) | BrainOOD on ABIDE · ABIDE | overall accuracy: 63.95 ± 4.65 -> 61.15 ± 2.84 | Hurt |
Surprising findings
- Most graph OOD methods (Mixup, DIR, GSAT, GMT) perform no better than ERM or simple GNN baselines on brain networks.
- Existing graph OOD methods with a GCN backbone fail to improve over the plain GCN baseline.
Limitations
- Reported OOD gains have large standard deviations, so differences between methods are often within noise.
- Evaluation uses only two brain network datasets (ABIDE and ADNI) with Schaefer-100 fMRI connectivity.
Experimental setup
Datasets: ABIDE ADNI
Benchmarks: ABIDE ADNI
Metrics: accuracy precision recall micro-F1 ROC-AUC
Baselines: ERM Deep Coral IRM GroupDRO VREx Mixup DIR GSAT GMT GCN GIN GAT
Builds on: GSAT GIB DIR GMT GOOD
Closest papers by meaning
- BrainMAP: Learning Multiple Activation Pathways in Brain Networks (AAAI 2025 · similarity 0.87)
- NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders (ICML 2025 · similarity 0.85)
- Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.85)
- Benchmarking Out-of-Distribution Generalization Capabilities of DNN-based Encoding Models for the Ventral Visual Cortex. (NeurIPS 2024 · similarity 0.84)
- Neurodegenerative Brain Network Classification via Adaptive Diffusion with Temporal Regularization (ICML 2024 · similarity 0.84)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.