Graph Neural Network Learning on the Pediatric Structural Connectome

Motivation: The structural connectome is a naturally occurring brain connectivity graph useful for studying cognitive function, yet machine learning applications on the connectome remain largely unexplored in pediatric populations. Goal(s): We aimed to train models for pediatric connectome sex classification, a clinically relevant benchmark for learning on the connectome. Approach: We trained two graph neural networks (GNNs) and a multilayer perceptron (MLP) using data obtained from 135 pediatric patients. Pediatric data was enriched with connectomes from 309 adults to test the effect on model performance. Results: Enriching the pediatric dataset with adult data improved model performance. The best GNN achieved 84.4% pediatric classification accuracy. Impact: Our demonstrated 84.4% accuracy using GNNs to predict sex from pediatric structural connectomes underscored the capacity of GNNs to advance our understanding of sex-specific neurological development and highlighted the potential benefit of using adult connectomic data to enrich pediatric datasets.