WiDS Datathon 2025 – ADHD & Sex Prediction from fMRI and Metadata

This project was developed for the 2025 WiDS Datathon and for my CS 211 Final Project, focusing on predicting ADHD diagnosis and sex from fMRI functional connectome data and socio-demographic, emotional, and parenting metadata.

🧠 Overview

I implemented machine learning models to classify:

• ADHD diagnosis (binary)
• Sex (binary)

I tried out multiple machine learning models:

• XGBoost
• Fully Connected Neural Networks
• Graph Convolutional Neural Network (on connectomes only)

Data is from the Healthy Brain Network (HBN) dataset, provided by the datathon organizers. It includes:

• Functional brain connectivity matrices (fMRI)
• Socio-demographic and behavioral metadata

Download raw data provided:

kaggle competitions download -c widsdatathon2025

⚙️ Requirements

• Python ≥ 3.9
• NumPy, Pandas
• Scikit-learn, XGBoost
• PyTorch, PyTorch Geometric

Install dependencies:

pip install -r requirements.txt

📁 Project Structure

.
├── data/                 # Processed datasets 
├── models/               # GNN model 
├── GNN_helpers.py        # Core GCN implementation and training loop
├── utils/                # Preprocessing, metrics, and plotting
├── main.py               # Easy script to load data and train / test model
└── README.md

• Easy training and testing for GCN model with class and method definitions
• Includes specialized model GCN and flexible model XGboost for comparisons

To Dos

• Add MLP layer to GCN to include metadata in the model
• Clean up XGboost pipeline