Uncertainty-Aware Hypergraph Representation Learning for Neurodegenerative Disease Detection

Hypergraph-based representations provide a systematic way to model higher-order interactions among multiple brain regions, which is important for detecting neurodegenerative diseases. However, existing hypergraph learning methods for brain network analysis face two major challenges. First, conventional binarization of functional connectivity neglects intrinsic uncertainty induced by progressive and heterogeneous pathological processes, which can bias network construction. Second, under the limited neuroimaging data, generic contrastive augmentation strategies introduce biologically implausible perturbations, hindering effective representation learning. We propose an Uncertainty-aware Hypergraph Representation Learning framework (UH-FRL) for neurodegenerative disease detection. We first develop an Uncertainty Relationship Construction (URC) module to infer functional connectivity states under uncertainty by leveraging fuzzy set theory and fuzzy inference. To mitigate data scarcity, we introduce a fuzzy contrastive strategy that generates semantically consistent augmented samples guided by the inferred fuzzy connectivity, reducing unrealistic distortions. We further design a Fuzzy Hypergraph Encoder (FHE) that aggregates uncertainty-aware information from hyperedges and node features through fuzzy rule-based modeling and hypergraph message passing, enabling the learning of robust and discriminative representations of brain network organization patterns. We evaluate UH-FRL on two public datasets and one private clinical dataset. The results show that UH-FRL consistently outperforms state-of-the-art methods, demonstrating strong effectiveness and robustness for neurodegenerative disease detection. The code of implementation is available at: https://github.com/yincan-y/UH-FRL/tree/master.