Leveraging unified multi-view hypergraph learning for neurodevelopmental disorders diagnosis

Accurate diagnosis of neurodevelopmental disorders relies on understanding the complex interactions and high-order relationships between brain regions. This work aims to model the subtle, disease-specific high-order relationships among brain regions that have been overlooked in existing works.

This paper proposes a Unified Multi-View Hypergraph Learning framework that combines knowledge-driven and data-driven strategies for a more precise and comprehensive representation of the adolescent brain network. The knowledge-driven branch leverages prior knowledge of functional brain subnetworks to guide feature learning and uncover structured, high-order functional associations. Meanwhile, the data-driven branch consists of two complementary modules: at the global level, a nearest-neighbor-based strategy captures large-scale associations involving overlapping brain regions; at the local level, a granularity-adaptive approach identifies finer, region-specific high-order relationships, allowing for a more nuanced understanding of brain network interactions.

Experimental results on the ABIDE and ADHD datasets demonstrate that our method outperforms existing methods in diagnostic accuracy and robustness. Additionally, visualizing the high-order associations learned from both branches reveals new insights into the pathogenic mechanisms of these disorders.

The proposed method combines knowledge-driven and data-driven strategies for high-order brain network modeling, advancing the understanding of brain networks in neurodevelopmental diseases.