Multi-atlas brain network enables richer representations by integrating complementary parcellation schemes. However, existing multi-atlas fusion methods overlook spatial overlap between atlases, leading to redundant encoding and entangled representations that amplify redundant information while suppressing discriminative features. This limits the rich representational advantages of multi-atlas methods. To address this challenge, we propose a novel Disentangled Multi-Atlas High-Order Representation Learning Network (DMH-Net) to enable more discriminative fusion of multi-atlas features. Specifically, the Multi-Atlas High-Order Brain Network (MA-HBN) module is proposed to construct high-order brain networks, effectively enhancing the representation of complex brain interaction patterns. Subsequently, we design a Contrastive Learning Decoupling (CLD) module that leverages contrastive learning theory to constrain the similarity between common and specific features across atlases, thereby removing cross-atlas redundancy while reinforcing inter-atlas complementarity, which significantly improves the discriminative power of multi-atlas fusion. Finally, we employ a feature enhancement strategy to ensure that the decoupled features are more discriminative. Experimental results on two real-world datasets demonstrate that our method effectively exploits the representational richness of multi-atlas and achieves substantial performance improvements.