Brain Network Analysis — Machine Learning Pipeline

This project analyzes brain network connectivity data using classical machine learning models and graph neural networks. Each subject is represented by a structural brain connectivity matrix stored as a .mat file, along with subject metadata stored in an Excel file. The pipeline extracts edge-level, graph-theoretic, spectral, and hemispheric asymmetry features, then evaluates multiple predictive models using nested cross-validation.

Project Goal

The goal of this project is to predict subject-level attributes from brain connectivity patterns. The notebook evaluates three classification tasks:

1. Sex classification

   • Female vs. Male

2. Creativity classification

   • Normal vs. Creative

3. Mathematical ability classification

   • Normal vs. High Math

The project compares different feature representations and machine learning models to determine which combination performs best for each task.

Dataset

The notebook expects the following structure:

brainnetworks/
├── smallgraphs/
│   ├── subject_1_fiber.mat
│   ├── subject_2_fiber.mat
│   └── ...
├── metainfo.xls
└── brain_network_analysis_v5.ipynb

Each .mat file should contain a variable named:

fibergraph

The notebook is currently configured for 70 × 70 brain connectivity matrices.

The metadata file should contain at least the following columns:

Column	Description
URSI	Subject identifier used to match metadata with .mat files
Sex	Subject sex label
Subject_type	Group label used for creativity and math classification

In the current run, the notebook loaded 114 matched subjects after filtering metadata rows against the available .mat files.

Feature Engineering

The pipeline builds several feature sets from each brain connectivity matrix.

1. Edge Features

The upper triangle of each 70 × 70 matrix is extracted to avoid duplicate symmetric edges.

Edge features: 2,415

2. Graph and Spectral Features

The notebook computes graph-level statistics using NetworkX, including centrality and connectivity-related summaries. It also extracts spectral features from the normalized graph Laplacian.

Graph + spectral statistics: 63

3. Hemispheric Asymmetry Features

The pipeline computes left-right hemispheric asymmetry features.

Asymmetry features: 10

4. Combined Feature Set

All feature groups are concatenated into one representation.

Total combined features: 2,488

Models Used

The notebook evaluates the following models:

• Support Vector Machine with mutual information feature selection
• Support Vector Machine with PCA
• L1-regularized Logistic Regression
• Random Forest
• Extra Trees
• Gradient Boosting
• Multi-Layer Perceptron
• XGBoost
• LightGBM
• Graph Convolutional Network
• Graph Isomorphism Network

The GNN models are only executed if PyTorch and PyTorch Geometric are installed.

Evaluation Method

The project uses nested stratified cross-validation:

Outer folds: 5
Inner folds: 3
Random seed: 42

The inner loop performs hyperparameter tuning with GridSearchCV, while the outer loop estimates generalization performance.

The main evaluation metrics are:

• Accuracy
• Balanced accuracy
• F1-score
• ROC-AUC
• Average precision

For imbalanced tasks, the pipeline uses class balancing techniques such as BorderlineSMOTE when imbalanced-learn is available. It also applies threshold tuning using Youden's J statistic.

Results

Best models from the current notebook run:

Task	Best Model	Feature Set	Accuracy	Balanced Accuracy	F1
Sex classification	Gradient Boosting	Edge upper triangle	0.693	0.676	0.748
Creativity classification	Gradient Boosting	Topology	0.723	0.519	0.071
High math classification	Extra Trees	Edge upper triangle	0.793	0.550	0.182

Although creativity and high-math classification achieved relatively high accuracy, their balanced accuracy and F1-scores show that the class imbalance makes these tasks harder. Accuracy alone should not be treated as the final indicator of model quality.

Installation

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate

For Windows:

python -m venv .venv
.venv\Scripts\activate

Install the required packages:

pip install numpy pandas scipy scikit-learn matplotlib networkx statsmodels xlrd jupyter

Optional packages for additional models:

pip install imbalanced-learn xgboost lightgbm torch

PyTorch Geometric installation depends on the installed PyTorch and CUDA version. Follow the official PyTorch Geometric installation instructions if GNN support is required.

Running the Notebook

1. Open the notebook:

jupyter notebook brain_network_analysis_v5.ipynb

2. Update the dataset paths in the configuration cell:

MAT_DIR = "path/to/smallgraphs"
META_PATH = "path/to/metainfo.xls"
OUTPUT_DIR = "path/to/output_directory"

3. Run all notebook cells from top to bottom.

The notebook will:

• Load the brain connectivity matrices
• Match .mat files with subject metadata
• Normalize the connectivity graphs
• Extract edge, topology, spectral, and asymmetry features
• Build classification tasks
• Train and evaluate multiple models
• Print task-wise and overall best results

Project Structure

.
├── brain_network_analysis_v5.ipynb
├── README.md
└── data/
    ├── smallgraphs/
    │   └── *.mat
    └── metainfo.xls

Notes

• The notebook expects each .mat file to contain a fibergraph variable.
• The current configuration assumes each brain network is a 70 × 70 matrix.
• If your matrices use a different number of nodes, update the N_NODES value in the configuration cell.
• XGBoost, LightGBM, imbalanced-learn, PyTorch, and PyTorch Geometric are optional. The notebook will skip or warn about unavailable optional libraries.
• Results may vary depending on available subject files, metadata filtering, package versions, and hardware.

Summary

This project demonstrates an end-to-end brain network classification pipeline using structural connectivity matrices. It combines graph-based feature engineering, classical machine learning, imbalance-aware evaluation, and optional graph neural networks to study whether brain connectivity patterns can predict subject attributes such as sex, creativity group, and high mathematical ability.