Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Exploring Structural Connectivity Networks for Classification of Post-Stroke Aphasia Patients and Healthy Controls using Graph Neural Networks
Poster A51 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Samaneh Nemati1, Roger Newman-Norlund2, Natalie Hetherington1, Leo Bonilha3, Julius Fridriksson1; 1University of South Carolina, Department of Communication Sciences and Disorders, 2University of South Carolina, Department of Psychology, 3Emory University, Department of Neurology
Background: Structural connectivity, which is derived from anatomical connections (white matter fiber tracts) connecting cortical and subcortical brain regions, provides a way to construct brain graphs and investigate the organization of brain networks in patients with post-stroke aphasia (PSA). Graph-based features, derived from these structural connectivity matrices, can be used to train machine learning classifiers, and automatically diagnose brain disorders. Graph neural networks (GNNs) are an extension of traditional convolutional neural networks that are designed to operate on graph-structured data. It leverages the structural information present in the graph to perform tasks such as node classification, link prediction, and graph classification. Method: In this study, the deep graph library (DGL) was used to implement and apply GNNs for classifying patients with chronic post-stroke aphasia (N=50) compared to age-matched healthy controls (HC, N=40). The input to GNN model was the adjacency matrices (created based on structural connectivity) and a feature matrix, which represents the features associated with each node (brain regions). Nodes of the adjacency matrix of each individual were defined by parcellating the brain using the JHU atlas. Edges of the adjacency matrix were defined by calculating the structural coupling between each pair of nodes. To define the feature matrix, we used time-series recorded during resting-state functional magnetic resonance imaging (rsfMRI) scans. The data were structured similarly for both healthy and PSA groups. GNN was developed using DGL toolbox (developed in Python). The architecture includes six consecutive Graph Convolutional layers followed by Batch Norm and ReLU activation functions. At the end of feature decoding path, a linear function maps the resulting features into number of classes defined in this study. Using such a model architecture, in each layer, a node's representation is updated by considering both its own features and the features of its neighbors. By stacking multiple layers, the network can capture increasingly complex patterns and dependencies in the graph structure. Results: The model was validated using three performance metrics: precision, recall, and F1 score. Preliminary results showed that classifying PSA and healthy controls using GNN provides 0.85, 0.77, 0.76 for precision, recall, and F1 score, respectively. Conclusion: GNNs can learn patterns and relationships within the brain networks and make predictions about the presence or absence of PSA. In the next step, we will use this trained model to explore structural brain connections supporting behavioral performance of patients with PSA in language-related tasks.
Topic Areas: Disorders: Acquired, Computational Approaches