Predicting cognitive scores with graph neural networks through sample selection learning.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Predicting intelligence quotient (IQ) from functional brain connectomes is crucial for understanding brain function in health and disease.
• Existing machine learning models often overlook the topological properties of brain connectomes by vectorizing them.

Purpose of the Study:

• To develop a novel Graph Neural Network (GNN) model, RegGNN, for predicting IQ scores from brain connectivity that preserves topological information.
• To introduce an effective sample selection method to enhance the performance of deep learning models for IQ prediction.

Main Methods:

• Designed a novel regression GNN (RegGNN) model to predict IQ scores from brain connectivity, leveraging graph-based topological features.
• Developed a learning-based sample selection method utilizing the properties of symmetric positive definite (SPD) matrices to identify informative training samples.
• Employed 3-fold cross-validation for evaluating prediction performance.

Main Results:

• RegGNN model achieved superior performance in predicting full-scale and verbal IQ scores in autism spectrum disorder (ASD) cohorts compared to existing methods.
• The proposed sample selection method demonstrated generalizability, improving performance for other learning-based approaches.
• Competitive IQ prediction performance was achieved for neurotypical subjects.

Conclusions:

• The RegGNN model effectively predicts IQ scores by incorporating brain connectome topology, offering advancements over traditional vectorization methods.
• The novel sample selection strategy enhances the efficiency and accuracy of deep learning models for cognitive prediction tasks.
• This approach holds promise for advancing our understanding of intelligence and brain connectivity in both clinical and healthy populations.