We Need to Talk About Functional Brain Networks

This repository contains implementations of various machine learning models (GCN, GAT, Transformer, MLP, and DeepSet) applied to open fNET datasets (ABIDE and HCP-Gender) for brain network analysis.

Dataset Description

ABIDE Dataset

• Task: Autism Spectrum Disorder (ASD) vs Control classification
• Features: ROI-based functional connectivity matrices
• Atlas: Schaefer 400-parcel atlas
• Sample Size: ~800 subjects

HCP-Gender Dataset

• Task: Gender classification from resting-state fMRI
• Features: ROI-based functional connectivity matrices
• Atlas: Schaefer 1000-parcel atlas
• Sample Size: ~1,300 subjects

Model Architectures

• MLP: Multi-Layer Perceptron on flattened connectivity matrices
• DeepSet: Permutation-invariant architecture for set-structured data
• Transformer: Self-attention based model for graph-level prediction
• GNN: Graph Neural Networks (GCN/GAT) operating on brain connectivity graphs

Usage

Each Python file runs the corresponding model on its dataset with 5 different random seeds for robust evaluation:

ABIDE Dataset

# Run MLP model on ABIDE
cd ABIDE
python ABIDE-MLP.py

# Run DeepSet model on ABIDE  
python ABIDE-DeepSet.py

# Run Transformer model on ABIDE
python ABIDE-Transformer.py

# Run GNN models on ABIDE
python ABIDE-GNN.py

HCP Dataset

# Run MLP model on HCP-Gender
cd HCP-Gender
python HCP-Gender-MLP.py

# Run DeepSet model on HCP-Gender
python HCP-Gender-DeepSet.py

# Run Transformer model on HCP-Gender
python HCP-Gender-Transformer.py

# Run GNN models on HCP-Gender
python HCP-Gender-GNN.py

Evaluation Protocol

• 5 random seeds for statistical reliability
• Stratified train/val/test split (70%/10%/20%)
• Metrics: Accuracy, F1-score, AUC-ROC
• Results reported: Mean ± standard deviation across seeds