Predicting Cognitive Scores With Graph Neural Networks Through Sample Selection Learning

Abstract

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

Meta-RegGNN: Predicting Verbal and Full-Scale Intelligence Scores Using Graph Neural Networks and Meta-learning

“…Evaluation and comparison method. To benchmark our method, we chose the first and unique deep learning method proposed in the literature that uses GNN to predict cognitive scores [7] without the proposed sample selection step. The results for the ASD and NT cohorts for FIQ and VIQ are shown in Fig.…”

“…Compute adapted parameters Θ i according to Equation 2 8: Update Θ according to Equation 3 using LT i in Equation 19: end networks [23], all negative eigenvalues are set to zero to train our regression GNN [7]. Thus, regression GNN receives the regularized positive adjacency matrix I of a connectome and predicts the corresponding behavioral scores using graph convolutions.…”

“…Intelligence, in particular, is a key aspect of ASD. State-of-the-art methods successfully used functional brain connectomes to predict cognitive measures such as Intelligence Quotient (IQ) scores in both disordered and healthy cohorts [6,7,8]. Indeed, functional brain connectomes describe the brain network structure and are derived from resting-state magnetic resonance imaging (MRI).…”

“…However, a lack of works that explored GNN for the prediction of cognitive scores has been noticed. Hanik et al [7] was the first to propose a GNN architecture, called RegGNN, specialized in regressing brain connectomes to a cognitive score to predict. To better improve the performance of GNN, they also proposed a learning-based sample selection method that selects training samples with the highest predictive power.…”

Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome

“…Since brain networks can be treated as graphs, one of the most popular architectures for building such prediction models is Graph Convolution Networks (GCN) [7]. However, the majority of GCN-based studies on brain connectivity focus on a single imaging modality [9,4,5] but fall short in leveraging the relationship between FC and SC. Existing works that perform multi-modal fusion [21] either use two GCNs to extract features from SC and FC graphs separately [23] or directly discard the graph structure from one modality (e.g., regarding FC as features defined on the SC graph) [10].…”