Subject-Invariant Domain Generalization for Psychiatric Disorder Identification

ICLR 2026 Conference Submission18406 Authors

Keywords: psychiatric disorder identification, brain network analysis

Abstract: Analyzing functional brain networks has emerged as a critical approach for understanding and diagnosing psychiatric disorders. Existing approaches primarily follow the standard supervised learning, which assumes that source and target data are independent and identically distributed. However, due to substantial inter-subject distributional differences in brain network data, models built on this assumption struggle to generalize from source to target datasets, resulting in suboptimal diagnostic performance. To address this issue, we propose a two-stage Subject-Invariant Domain Generalization (SIDG) model that learns subject-invariant representations in the pre-training stage, enabling their effective use for better psychiatric disorder identifcation in the fine-tuning stage. In order to overcome the mismatch between single-level topological representation methods and the inherently hierarchical topology of brain networks, we introduce a novel Hierarchical Topology Enhanced Graph Transformer Reconstruction (HTE-GTR) module to thoroughly learn subject-invariant representations distributed across multiple topological levels. Furthermore, we design tailored Subject-Invariant Reconstruction (SIR) loss comprising a subject-invariant term and a reconstruction term, to mitigate the impact of inter-subject distributional differences while preserving discriminative information for downstream tasks. Experiment results show clear improvements of our proposed SIDG on both the public ABIDE and ADHD datasets. The code is available at https://anonymous.4open.science/r/SIDG.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 18406