Diabetes is a common major chronic disease characterized by high incidence rates and complex complications. This paper constructs a Diabetes Knowledge Graph (DKG). At the schema layer, DKG defines 17 entity types and 85 relation types centered around diseases. At the data layer, we develop a verification-enhanced large language model framework for the medical field (MED-VerifyKG). This framework integrates LLM-based candidate triple extraction, rule-based filtering, and expert validation, with iterative prompt and rule refinement. Triple fusion is ultimately prioritized based on expert validation confidence. The constructed DKG contains 18,720 entities and 24,249 triples, providing robust data support for diabetes-related intelligent question answering and clinical decision support.