Diabetes is a prevalent chronic disease, and early screening is critical for prevention and timely intervention. While traditional machine learning models treat individuals independently in tabular survey data such as NHANES, we investigate whether graph neural networks can exploit the latent patient-similarity structure for diabetes risk prediction. We construct a multi-relational patient–patient graph from NHANES, where each relation is derived from a distinct modality (diet, examination, laboratories, medications, and questionnaires). Building on MedGCN, we further study multi-relational learning and relation-level attention to better leverage heterogeneous relations. Experiments show that explicitly modeling multiple relations improves performance over single-relation baselines; relation-level attention yields stronger discrimination (ROC-AUC and F1), while R-GCN achieves slightly higher minority-class recall under class imbalance.