HOI-Brain: a novel multi-channel transformers framework for brain disorder diagnosis by accurately extracting signed higher-order interactions from fMRI

A research team led by Dengyi Zhao has introduced HOI-Brain, a groundbreaking computational framework published in Medical Image Analysis. The system uses a novel multi-channel transformer architecture to analyze functional Magnetic Resonance Imaging (fMRI) data, moving beyond traditional graph-based models that only capture simple pairwise connections. Its core innovation is the accurate extraction of 'signed higher-order interactions'—complex patterns of communication involving multiple brain regions that can be either positively or negatively synergistic. This is achieved through a co-fluctuation measure based on Multiplication of Temporal Derivatives (MTD), which provides temporal resolution.

HOI-Brain encodes these interactions into signed weighted simplicial complexes and applies two filtration processes from Persistent Homology theory to extract spatiotemporal neural organizations. The integrated multi-channel transformer then synthesizes these heterogeneous topological features for diagnosis. In experiments across three major disorder datasets—Alzheimer's disease, Parkinson's syndrome, and autism spectrum disorder—the framework demonstrated superior diagnostic performance, effectiveness, and interpretability compared to existing methods. Critically, the key brain regions and higher-order patterns it identified align with established neuroscience literature, offering meaningful biological insights rather than just a black-box prediction.