To address the challenges of entity dilution in single samples and the underutilization of complete entity location information in traditional named entity recognition (NER) methods based on reading comprehension frameworks, this study proposes a medical named entity recognition model that integrates key features of entity start and end positions (IKFSE). First, a key feature extraction module is designed to capture important features of the start and end positions of medical entities, thus reducing the impact of redundant information on the model. Second, an entity head-tail feature fusion module is introduced, which incorporates the mutual influence between the start and end positions during entity prediction. This integration enhances the model’s ability to represent semantics by utilizing full positional information. Experimental results on two public datasets, cEHRNER and CCKS2017, demonstrate that the proposed method outperforms several mainstream sequence labeling models and reading comprehension models in Chinese medical NER tasks.