brain-network-core

---
name: brain-network-core
description: CORE framework for out-of-distribution generalization in brain network analysis via site-aware confounder decoupling
category: neuroscience
created: "2026-05-09"
readiness_status: available
dependencies: []
---

# CORE: Confounded Representation Elimination for Brain Network Analysis

## Overview

This methodology presents the CORE framework for out-of-distribution (OOD) generalization in brain network analysis. Published on arXiv (2605.06050v1, May 2026), it addresses the critical challenge of site-specific confounders that degrade model performance when brain network models are deployed across different scanning sites, populations, or protocols.

## Key Innovation

### Site-Aware Confounder Decoupling
- Identifies and separates site-specific artifacts from biologically meaningful brain network patterns
- Uses a disentanglement approach that explicitly models the confounding structure
- Enables models to learn site-invariant representations while preserving diagnostic signal

### Prior-Guided Disentanglement
- Incorporates neuroscientific priors to guide the separation of confounded vs. clean representations
- Uses structural information about brain networks to constrain the disentanglement process
- Ensures that biologically implausible patterns are not learned as "confounders"

### OOD Risk Formulation
- Formalizes the problem as minimizing worst-case risk across unseen sites
- Uses a distributionally robust optimization approach
- Provides theoretical guarantees on generalization bounds

## Architecture

### Input Representation
- **Brain Networks**: Functional connectivity matrices (FC) from fMRI data
- **Site Labels**: Metadata indicating acquisition site/scanner/protocol
- **Graph Structure**: Nodes represent brain regions, edges represent functional connectivity

### CORE Framework Components

1. **Confounded Encoder**: Maps brain networks to a latent representation that may contain site-specific artifacts
2. **Site Predictor**: Attempts to predict site labels from the latent representation
3. **Disentanglement Module**: Separates the representation into:
   - Site-invariant component (biologically meaningful)
   - Site-specific component (confounders)
4. **Classifier**: Makes predictions using only the site-invariant representation
5. **Prior-Guided Regularizer**: Ensures biological plausibility of the disentanglement

### Training Objective
- Minimize classification loss on site-invariant features
- Minimize site predictability from invariant features (adversarial objective)
- Maximize site predictability from site-specific features
- Apply prior-guided regularization to maintain biological consistency

## Implementation Details

### PyTorch Implementation
- **Backbone**: 2-layer Graph Convolutional Network (GCN)
- **Hidden Size**: 64 dimensions
- **Optimizer**: Adam with learning rate 1e-3, weight decay 5e-4
- **Batch Size**: Configurable based on dataset size
- **Epochs**: Typically 100-200 with early stopping

### Dataset Requirements
- Multi-site fMRI datasets with site labels
- Examples: ABIDE, ADHD-200, UK Biobank, HCP
- Preprocessed functional connectivity matrices (e.g., using atlas-based parcellation)

### Baselines for Comparison
- Standard GCN without confounder handling
- Domain adversarial neural networks (DANN)
- Invariant risk minimization (IRM)
- Site-specific fine-tuning

## Application Workflow

1. **Data Preparation**: Collect multi-site brain network data with site labels
2. **Preprocessing**: Compute functional connectivity matrices using standard pipelines
3. **Model Training**: Train CORE framework with confounder decoupling
4. **Validation**: Evaluate OOD generalization on held-out sites
5. **Deployment**: Use site-invariant model for cross-site predictions

## Key Advantages

- **Robustness**: Maintains performance across different scanning sites and protocols
- **Interpretability**: Disentangled representations provide insights into confounding structure
- **Theoretical Guarantees**: Formal bounds on OOD generalization performance
- **Flexibility**: Applicable to various brain network analysis tasks (classification, regression, etc.)

## When to Use

- Multi-site fMRI studies where data comes from different scanners/protocols
- Brain disorder diagnostic models that need to generalize across populations
- Research into confounder-aware machine learning for neuroscience
- Building robust biomarkers that are not site-specific
- Domain generalization in neuroimaging applications

## Pitfalls

- Requires multi-site data with site labels for training
- Disentanglement quality depends on the strength of site-specific confounders
- May reduce in-distribution performance slightly in exchange for OOD robustness
- Prior selection is critical — incorrect priors can harm disentanglement
- Computational overhead from adversarial training components
- Not suitable for single-site studies without domain shift concerns

## References

- arXiv: 2605.06050v1 — "When Brain Networks Travel: Learning Beyond Site"
- Domain generalization literature: DANN, IRM, CORAL
- Brain network analysis: functional connectivity, graph neural networks
- Confounder handling: causal inference, disentanglement learning