NeuroGraph: Benchmarks for Graph Machine Learning in Brain Connectomics.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Machine learning effectively analyzes high-dimensional functional neuroimaging data for predicting neurological and psychiatric conditions.
• Graph-based representations are crucial for modeling brain region interactions in functional MRI research.
• Applying graph machine learning to neuroimaging is challenging due to extensive preprocessing pipelines and parameter spaces.

Purpose of the Study:

• Introduce NeuroGraph, a collection of graph-based neuroimaging datasets.
• Demonstrate the utility of NeuroGraph for predicting behavioral and cognitive traits.
• Provide open-source tools to advance graph-based neuroimaging analysis.

Main Methods:

• Crafted 35 datasets covering static and dynamic brain connectivity.
• Benchmarked over 15 baseline methods for graph machine learning on neuroimaging data.
• Developed generic frameworks for learning on static and dynamic graphs.

Main Results:

• Correlation vectors as node features, larger numbers of regions of interest, and sparser graphs improved prediction performance.
• Identified key parameters influencing the effectiveness of graph-based neuroimaging analysis.
• Established baseline performance across various graph construction and machine learning methods.

Conclusions:

• NeuroGraph facilitates robust graph-based analysis of functional neuroimaging data.
• The findings provide practical guidelines for optimizing graph construction in neuroimaging studies.
• The open-source Python package supports reproducible research and further development in the field.