This study presents a high-accuracy framework for medical entity-relation extraction, which is built on Qwen-based feature reading. To tackle the issue of error propagation that occurs between entity extraction and relation classification in medical texts, a data augmentation approach using pseudo-entities is proposed. First, the Transformer‑based Qwen model, enhanced with a feature‑reading unit, improves recognition of long entities and complex relations in medical texts, thereby increasing entity‑extraction accuracy. Second, within a traditional pipeline extraction framework, a relation negative‑example generation module is added; combined with a pseudo‑entity generation model, three data‑augmentation strategies effectively reduce the negative impact of entity‑extraction errors on relation classification. Finally, to cope with the additional training time caused by data augmentation, a floating‑label strategy is adopted, significantly improving training efficiency and lowering computational costs. Experiments on the CDR dataset show that the proposed model raises entity‑extraction accuracy by 1.18% over the runner‑up model PL‑Marker, and in entity‑relation extraction improves the F1 value by 3.07% and recall by 6.57% compared with the runner‑up model CBLUE, effectively mitigating the negative influence of entity‑extraction errors on relation classification in traditional pipeline models.