This repository contains the PyTorch implementation used for the revised manuscript. The code is organized to reproduce the main cross-validation experiments with fMRI functional connectivity and DTI structural connectivity matrices.

The proposed model is implemented in PyTorch. The default configuration follows the revised manuscript:

• Optimizer: Adam
• Initial learning rate: 5e-5
• Weight decay: 5e-6
• Epochs: 300
• Batch size: 20
• Cross-validation: 10 folds with random seed 7
• Local branch: single-layer graph attention network integrated with self-attention
• Global branch: PageRank-based graph attention integrated with self-attention
• Classifier: flattened global-local features followed by a multi-layer perceptron

config/ Experiment configuration
data/ Data loader and data-format documentation
models/ GLGAN model implementation
training/ Cross-validation training and metrics
utils/ Logging and reproducibility helpers
experiments/ Reproduction commands for main and ablation experiments
analysis/ ROC, t-SNE, parameter-count, hubness, sensitivity, and ROI analysis scripts
tests/ Basic checks for data loading, model forward pass, and training
main.py Command-line entry point

conda create -n glgan python=3.8
conda activate glgan
pip install -r requirements.txt

The experiments in the manuscript were run on an NVIDIA GeForce RTX 3080 12GB GPU. CPU execution is supported for installation checks but is not recommended for full 300-epoch cross-validation.

ADNI data are governed by the ADNI data-use agreement and are not redistributed in this repository. After obtaining access from ADNI, place the processed .mat files under:

data/raw/
|-- ADNI_fmri.mat
`-- ADNI_DTI.mat

Expected arrays:

• fMRI connectivity: [n_subjects, 90, 90], or regional time series [n_subjects, 90, time_points] from which the loader computes Pearson connectivity.
• DTI connectivity: [n_subjects, 90, 90].
• Labels: a vector stored in the fMRI .mat file. Use --label-key if the key is not one of label, labels, y, Y, gnd, or target.

See data/README.md for details.

With the default file names:

python main.py

If your .mat variable names differ:

python main.py --fmri-key fmri --dti-key dti --label-key label

To select a binary task from multi-class labels:

python main.py --class-values 0 1 --output-dir results/NC_SMC

Results are written to:

results/
|-- fold_metrics.csv
|-- aggregate_metrics.csv
|-- training.log
`-- checkpoints/

The code exposes the main ablation switches used to reproduce component-removal experiments:

python main.py --disable-local --output-dir results/ablation_no_local
python main.py --disable-global --output-dir results/ablation_no_global
python main.py --disable-pagerank --output-dir results/ablation_no_pagerank
python main.py --disable-self-attention --output-dir results/ablation_no_self_attention

Additional notes are in experiments/README.md.

Post-processing scripts for common manuscript analyses are provided under analysis/. They operate on outputs generated by the GLGAN training workflow.

python -m analysis.parameter_count --output results/analysis/parameter_count.csv
python -m analysis.roc_analysis --input predictions.csv --output results/analysis/roc_points.csv
python -m analysis.tsne_analysis --features features.npy --labels labels.npy --output results/analysis/tsne_coordinates.csv
python -m analysis.hubness_analysis --adjacency adjacency.npy --output results/analysis/hubness.csv
python -m analysis.roi_ranking --input results/analysis/hubness.csv --output results/analysis/top_rois.csv
python -m analysis.sensitivity_statistics --input sensitivity_results.csv --parameter-column learning_rate --metric-column accuracy

The ROC script expects a CSV with label and score columns. The t-SNE script expects a NumPy feature matrix shaped [n_samples, n_features].

To confirm that the repository is installed correctly:

python -m unittest discover -s tests -v

Please cite the corresponding manuscript if you use this code.