Abstract
Insomnia disorder (ID) is neurobiologically heterogeneous and often eludes characterization by traditional group-level neuroimaging. Subtyping based on neuroimaging and clinical data offers a promising strategy for identifying biologically and clinically meaningful ID subgroups. To address this need, we developed a Gray Matter Population Graph Attention Autoencoder (GM-PGAAE) to subtype insomnia disorder in a cohort comprising 140 patients diagnosed with ID and 57 matched healthy controls. Each subject was represented as a node defined by atlas-based gray matter (GM) volumes. Population edges combined imaging-derived intersubject correlations with clinical similarity via a Hadamard product, generating an adjacency matrix that jointly encodes structural and phenotypic relationships. A Graph Attention Autoencoder learned low-dimensional embeddings that adaptively weighted informative intersubject connections, and clustering these embeddings identified distinct subtypes. Regional and network-level differences were further assessed using Voxel-Based Morphometry (VBM) and individualized differential structural covariance networks (IDSCNs). Through this framework, two ID subtypes were identified. Compared with Subtype 2, Subtype 1 showed higher symptom severity and greater GM reductions-particularly in the cerebellar vermis, thalamus, middle occipital cortex, fusiform gyrus, and paracentral lobule-alongside negative associations between GM volume and clinical scores. IDSCNs further revealed reduced thalamocortical and subcortical Z-scores in Subtype 1, indicating subtype-specific network alterations. Overall, GM-PGAAE integrates structural MRI and clinical measures to derive individualized embeddings and delineate biologically distinct ID subtypes.