With the continuous advancement of research and technological progress, biomedical information is experiencing explosive growth. Relation extraction in the biomedical field holds significant research significance and practical value. However, in real-world scenarios, such as in healthcare, many relational facts often need to span multiple sentences to be fully expressed. Existing mainstream methods primarily employ sequential or graph models to represent entity relationships. However, these approaches often underutilize contextual information, leading to biases in entity relationship identification. To address this limitation, we introduce the MED-CDA model, namely, the Mention-Entity Double Fusion and Contrast Data Augmentation. This model comprises the Mention-Entity Double Fusion module and the Contrast data augmentation module. The Mention-Entity Double Fusion module utilizes a relation-specific mention attention network and a U-shaped document graph to model mention and entity information. Subsequently, by fusing the obtained features, it leverages contextual semantic information to fully capture the implicit logical relationships between entity pairs. The Contrast data augmentation module is a novel data augmentation method aimed at mitigating class imbalance issues in biomedical literature. This enables the model to focus more on text with fewer relation types, thereby enhancing its inference ability. Our experimental results exhibit substantial performance enhancements across three widely-used biomedical datasets—BIORED, CDR and GDA when compared to baseline models, underscoring its competitive edge.