The extraction of Chinese diabetes entity relations is the basis of Chinese diabetes medical data processing. However, the composition of Chinese diabetes medical data is complex and diverse, and there are problems such as entity sparsity and relationship overlap, which brings difficulties and challenges to the extraction of Chinese diabetes entity relations. The current mainstream entity relationship joint extraction model cannot make full use of text information, and the entity relationship extraction effect is poor. This study introduces a novel joint extraction approach that integrates multi-scale hybrid attention mechanisms with a hierarchical pointer network to achieve comprehensive feature representation and accurate relational triple prediction. The word vector of the text is obtained through the BERT, and context information of the word vector is obtained using BiLSTM to enhance the semantic information of the text vector; Multi-scale convolution kernels are introduced to capture information of different scales. The channel attention mechanism learns adaptive weights for feature channels, while the spatial attention mechanism computes position-wise importance weights, collectively enhancing the representation of discriminative features. Finally, the relationships between entities are identified through a hierarchical network. Experimental results demonstrate that the proposed method achieves superior performance on the DiaKG dataset, outperforming baseline approaches.