Biomedical entity-relation joint extraction constitutes a fundamental task for biomedical knowledge base development. As a crucial component in healthcare informatization systems, existing methods frequently exhibit performance deterioration when handling triple overlapping instances, primarily attributable to insufficient modeling of implicit entity-relation interactions. To compensate for the lack of Chinese biomedical datasets, the research creates a compact biomedical relationship extraction dataset by utilizing the medical records of 584 patients at a tertiary hospital. To overcome the challenge of effectively capturing implicit relationships, the research proposes BAGP, a Chinese biomedical joint extraction model that integrates MacBERT pre-training model, gated recurrent network with biaffine attention mechanism, and adversarial training. The evaluation outcomes reveal that F1-score metrics of 62.8% and 86.3% were achieved by the BAGP framework respectively on two benchmark datasets, namely the public CMeIE dataset and the proprietary CDeD dataset, both of which outperform the baseline model, providing a new paradigm for solving the complex entity-relationship interdependency problem.