EDGE-CONTRASTIVE PRETRAINED HETEROGENEOUS GRAPH NEURAL NETWORKS FOR MODELING ALZHEIMER’S DISEASE PROGRESSION

Abstract

We propose Edge-Contrastive Pretrained Heterogeneous Graph Neural Networks (ECP-HGNN), a new approach for analyzing Alzheimer’s disease (AD) progression with multi-modal patient data. Existing heterogeneous graph neural networks often struggle with missing or imbalanced edge data in clinical graphs, which limits their ability to capture nuanced patient relationships. ECP-HGNN tackles this issue by proposing a semantic-aware edge-contrastive pretraining approach, which directly learns resilient edge representations through comparisons between augmented views of distinct edge categories, including neuroimaging structural edges, blood biomarker edges, and clinical assessment edges. The pretrained edge embeddings are subsequently merged into a type-aware message passing layer, which supports dynamic aggregation of patient features without compromising clinical semantics. Additionally, the framework applies distinct encoders for genomic, proteomic, and clinical data, with a subsequent cross-modal fusion layer to predict AD progression scores. In contrast to traditional approaches, ECP-HGNN distinctively merges augmentations specific to edge types with contrastive learning, which guarantees that vital clinical relationships are preserved in pretraining. Experiments on extensive cohorts show the proposed method achieves greater prediction accuracy and robustness relative to current state-of-the-art approaches. The modular architecture of ECP-HGNN also supports scalable implementation across varied clinical datasets, which renders it an effective instrument for progressing precision medicine in neurodegenerative disorders.