Individual Functional Network Abnormalities Mapping via Graph Representation-Based Neural Architecture Search

Prenatal alcohol exposure (PAE) has garnered increasing attention due to its detrimental effects on both neonates and expectant mothers. Recent research indicates that spatio-temporal functional brain networks (FBNs), derived from functional magnetic resonance imaging (fMRI), have the potential to reveal changes in PAE and Non-dysmorphic PAE (Non-Dys PAE) groups compared with healthy controls. However, current deep learning approaches for decomposing the FBNs are still limited to hand-crafted neural network architectures, which may not lead to optimal performance in identifying FBNs that better reveal differences between PAE and healthy controls. In this paper, we utilize a novel graph representation-based neural architecture search (GR-NAS) model to optimize the inner cell architecture of recurrent neural network (RNN) for decomposing the spatio-temporal FBNs and identifying the neuroimaging biomarkers of subtypes of PAE. Our optimized RNN cells with the GR-NAS model revealed that the functional activation decreased from healthy controls to Non-Dys PAE then to PAE groups. Our model provides a novel computational tool for the diagnosis of PAE, and uncovers the brain’s functional mechanism in PAE.