Predicting cognitive scores with graph neural networks through sample selection learning

• DOI: 10.1007/s11682-021-00585-7
• Publication Date: 2021-11-10T06:02:36Z
• Authors: Martin Hanik; Mehmet Arif Demirtaş; Mohammed Amine Gh...; Islem Rekik
• Keywords: Neurotypical; Sample (material)

Analyzing the relation between intelligence and neural activity is of utmost importance in understanding working principles human brain health disease. In existing literature, functional connectomes have been used successfully to predict cognitive measures such as quotient (IQ) scores both healthy disordered cohorts using machine learning models. However, methods resort flattening connectome (i.e., graph) through vectorization which overlooks its topological properties. To address this limitation inspired from emerging graph networks (GNNs), we design a novel regression GNN model (namely RegGNN) for predicting IQ connectivity. On top that, introduce novel, fully modular sample selection method select best samples learn our target prediction task. since deep architectures are computationally expensive train, further propose learning-based that learns how choose training with highest expected predictive power on unseen samples. For this, capitalize fact their adjacency matrices) lie symmetric positive definite (SPD) matrix cone. Our results full-scale verbal outperforms comparison autism spectrum disorder achieves competitive performance neurotypical subjects 3-fold cross-validation. Furthermore, show approach generalizes other methods, shows usefulness beyond architecture.