Multimodal graph learning has been widely adopted for brain disease diagnosis by integrating diverse imaging modalities and modeling inter-subject relationships through population graph construction. However, existing methods typically face two critical limitations: (1) insufficient discriminative feature representations due to the redundancy across different imaging modalities, and (2) spurious graph connectivity introduced by nonlinear mappings in latent space-based graph construction. These issues hinder the effective utilization of multimodal information for diagnostic tasks. To overcome these challenges, we propose a novel multimodal disease prediction framework that enhances diagnostic performance through improved feature learning and graph construction. Specifically, we design a self-supervised contrastive learning approach to extract modality-aware representations, encouraging the model to emphasize modality-specific discriminative features. Additionally, to alleviate structural distortion arising from nonlinear feature space mappings, we incorporate a k-nearest neighbors (KNN) strategy to build initial subgraphs, which are used to complement the learnable population graph with robust and reliable topological priors. Extensive experiments on two publicly available datasets demonstrate that our proposed framework outperforms state-of-the-art baselines in disease prediction, achieving superior accuracy and robustness.