BrainConnect-ASD

BrainConnect-ASD is a research prototype for studying whether adversarial graph neural networks can reduce acquisition-site confounding in autism spectrum disorder (ASD) classification from resting-state fMRI. Its headline four-site LOSO evaluation achieves 0.7872 ROC AUC across 529 held-out subjects.

The project was developed for the AMD Developer Hackathon 2026 and trained on ABIDE I. It is intended for reproducible machine-learning research, not for clinical diagnosis, screening, or treatment decisions.

Research question

Multi-site neuroimaging models can learn scanner and acquisition artifacts instead of biological signal. BrainConnect-ASD combines a lightweight graph encoder with a gradient-reversal site classifier to encourage representations that are less informative about the acquisition site.

BOLD time series (T x 200 ROIs)
        |
        +-- sliding-window features
        +-- Fisher-z functional-connectivity graph
        |
        v
Adversarial Brain Mode Network
        |
        +-- ASD/typical-control research classifier
        +-- site-adversarial head (training only)

Key implementation details:

• CC200 functional atlas with 200 regions of interest
• 32 learnable graph modes
• approximately 105,000 trainable parameters
• gradient reversal for site-adversarial training
• leave-one-site-out (LOSO) evaluation

Results

The headline evaluation covers NYU, USM, UCLA, and UM. Its aggregate performance is shown first; the broader 20-site experiment is included as a separate generalization benchmark and powers the current multi-site demo.

Evaluation	Subjects	ROC AUC
Four-site LOSO evaluation (NYU, USM, UCLA, UM)	529	0.7872
Broader ABIDE I 20-site LOSO experiment	1,102	0.7298

The two rows describe different evaluation scopes and should not be pooled or treated as interchangeable. Each LOSO fold trains without the evaluation site. ABIDE II external evaluation is planned but has not yet been reported.

These results do not establish clinical validity. Performance can be affected by demographic imbalance, motion, preprocessing choices, site composition, and dataset shift.

Installation

Python 3.9 or newer is required.

git clone https://github.com/Yatsuiii/Brain-Connectivity-GCN.git
cd Brain-Connectivity-GCN
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
pytest -q

PyTorch installation can be platform-specific. If the default wheel is not suitable for your CPU, CUDA, or ROCm environment, install PyTorch using its platform instructions before installing this package.

Usage

Run inference on a preprocessed CC200 ROI time-series file:

python predict.py subject_rois_cc200.1D --verbose

Evaluate the published ensemble on ABIDE II after reviewing the download and dataset-use requirements:

python eval_abide2.py --n-subjects 50
python eval_abide2.py

Training and experiment entry points are documented in MODELS_AND_PIPELINE.md and EXPERIMENTS.md. The detailed research-use model card is in MODEL_CARD.md.

Repository layout

brain_gcn/             model, training, and evaluation modules
tests/                 unit and integration tests
hf_space/              Hugging Face demo source
results/               checked-in experiment summaries
predict.py             single-subject research inference
eval_abide2.py         external-dataset evaluation script
generate_finetune_data.py  synthetic explanation-data generator

Reproducibility and data

ABIDE data are not redistributed in this repository. Users are responsible for complying with the dataset's access and use terms. Random seeds, preprocessing choices, split definitions, checkpoint provenance, and exclusions should be reported alongside any derived result.

The generated LLM examples in finetune_data/ are synthetic templates. They are not clinician-authored reports, patient records, or evidence of clinical validation. They must not be used to infer subject-specific neurobiological findings without an independently validated attribution method.

Limitations

• ASD is heterogeneous and cannot be diagnosed from resting-state fMRI alone.
• ABIDE is a retrospective research dataset, not a prospective clinical cohort.
• Site-adversarial training does not prove removal of every scanner, motion, or demographic confound.
• The headline result has not yet been replicated on ABIDE II in this repo.
• Saliency and generated text are explanatory research aids, not causal or clinical interpretations.

Contributing

Issues and pull requests are welcome. Please read CONTRIBUTING.md and include tests and reproducibility details with behavioral changes.

Citation

Citation metadata are provided in CITATION.cff. Until a paper is available, cite the software repository and the exact release or commit used.

License

Licensed under the Apache License 2.0. Dataset and third-party model licenses remain separate.