Graph neural network based heterogeneous propagation scheme for classifying Alzheimer's disease

Abstract

Deep learning frameworks for disease classification using neuroimaging and non-imaging information require the capability of capturing individual features as well as associative information among subjects. Graphs represent the interactions among nodes, which contain the individual features, through the edges in order to incorporate the inter-relatedness among heterogeneous data. Previous graph-based approaches for disease classification have focused on the similarities among subjects by establishing customized functions or solely based on imaging features. The purpose of this paper is to propose a novel graph-based deep learning architecture for classifying Alzheimers disease (AD) by combining the resting-state functional magnetic resonance imaging and demographic measures without defining any study-specific function. We used the neuroimaging data from the ADNI and OASIS databases to test the robustness of our proposed model. We combined imaging-based and non-imaging information of individuals by categorizing them into distinctive nodes to construct a subject-demographic bipartite graph. The approximate personalized propagation of neural predictions, a recently developed graph neural network model, was used to classify the AD continuum from cognitively unimpaired individuals. The results showed that our model successfully captures the heterogeneous relations among subjects and improves the quality of classification when compared with other classical and deep learning models, thus outperforming the other models.