🧠 Graph4ASD Challenge

Live Leaderboard

📌 Overview

This repository hosts the Graph4ASD Challenge, a graph machine learning competition focused on Autism Spectrum Disorder (ASD) classification using resting-state fMRI functional connectivity data from the ABIDE dataset.

Participants must design and train Graph Neural Network (GNN) models to classify subjects as:

• ASD
• Typical Control (TC)

based on brain connectivity graphs.

Each subject is represented as a graph where:

• Nodes → brain regions (Craddock 200 atlas)
• Edges → functional connectivity between brain regions

---

🏆 Leaderboard

Leaderboard scores are automatically updated.

👉 Live Leaderboard

The evaluation metric is Macro F1-Score:

$$ \text{Macro F1} = \frac{F1_{ASD} + F1_{TC}}{2} $$

Rankings are sorted by descending score.

---

🧠 Task Description

Each sample corresponds to one subject’s brain graph.

Graph Definition

[ G = (A, X) ]

Where:

• A → adjacency matrix (functional connectivity)
• X → node feature matrix (identity features)
• Nodes → 200 brain regions (Craddock atlas)

Objective

Train a model that correctly predicts whether a subject:

• has Autism Spectrum Disorder (ASD)
• or is a Typical Control (TC)

using:

• Graph structure
• Node features

This is a graph-level classification task.

---

📊 Dataset Details

The dataset is derived from:

Autism Brain Imaging Data Exchange (ABIDE)

ABIDE aggregates anonymized rs-fMRI data from 17 international sites.

Full Dataset (original ABIDE)

Property	Value
Subjects	1,009
ASD	516 (51.14%)
Controls	493 (48.86%)
Brain atlas	Craddock 200

⚠️ For the challenge, a processed subset is provided, 484 samples for train and 153 for test. Participants must use data as-is, since original data has been exclusively processed for this challenge.

---

🧩 Data Representation

Each subject is stored as a graph:

• Adjacency matrix (A) → connectivity between brain regions
• Feature matrix (X) → node-level features

All graphs share:

• Same number of nodes → 200
• Same node ordering → Craddock atlas

---

📂 Dataset Access

The dataset is hosted in folder data/public.

📝 How to Submit Your Results

Step 1: Train

Train your model using:

• Graphs: Adjacency matrix and node feature matrix in data/public/adj_train.npy and data/public/node_features.npy. Each
• labels in data/public/train_label.csv

Step 2: Predict

Predict labels for every graph in:

• data/public/adj_test.npy and data/public/node_features_test.npy

Step 3: Prepare your submission files

Create a metadata.json that contains metadata about your submission:

{
  "team": "example_team",
  "run_id": "example_run_id",
  "type": "human",   // must be "human", "llm-only", or "human+llm"
  "model": "GAT",
  "notes": "Additional notes"
}

Create a CSV with columns id and y_pred (same format as data/public/sample_submission.csv):

id,y_pred
1,0
2,0
3,1
...

Encrypt your predictions.csv using: python extra/encrypt.py predictions.csv.

Then put predictions.csv.enc and metadata.json in the submissions/ folder.
Note: ❗ Do NOT upload the raw CSV. You need to submit an encrypted version of your predictions file to keep privacy.

Step 4: Create a Pull Request

Commit and push the predictions.csv.enc and metadata.json to the repository. Then, open Pull Request or push to the main branch. The automated pipeline will decrypt and score your submission to update the leaderboard.

⚠️ Only one submission per participant is allowed.

🏁 Challenge Rules

• You are free to use any Graph Neural Network.
• You are not allowed to use external data.
• You must not try to identify each subject.

📚 References

Cameron Craddock, Yassine Benhajali, Carlton Chu, Francois Chouinard, Alan Evans, András Jakab, Budhachandra Singh Khundrakpam, John David Lewis, Qingyang Li, Michael Milham, Chaogan Yan, Pierre Bellec (2013). The Neuro Bureau Preprocessing Initiative: open sharing of preprocessed neuroimaging data and derivatives. In Neuroinformatics 2013, Stockholm, Sweden.