Medical named entity recognition based on domain knowledge and position encoding

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• Accurate named entity recognition (NER) in Chinese electronic medical records (EMRs) is crucial for clinical data analysis.
• Existing models often struggle with precise boundary detection and incorporating domain-specific medical terminology.

Purpose of the Study:

• To develop a novel NER model for Chinese EMRs with enhanced boundary detection capabilities.
• To leverage medical domain knowledge and advanced positional encoding techniques to improve recognition accuracy.

Main Methods:

• A BERT module integrated with a lexical adapter for medical domain-specific terms.
• Feature extraction using Star-Transformer for local features and BiLSTM for long-distance features.
• Incorporation of Rotary Position Embedding (RoPE) to enhance semantic feature extraction.

Main Results:

• Achieved an F1-score of 85.78% on the CCKS2020 dataset, a 2.96% increase over the baseline.
• Demonstrated improved performance on a self-built breast cancer ultrasound report dataset.
• Validated the model's effectiveness and applicability in the medical field.

Conclusions:

• The proposed model significantly improves named entity recognition in Chinese EMRs.
• The integration of domain knowledge, advanced encoding, and specific architectures enhances boundary detection and overall accuracy.
• The model shows strong potential for real-world applications in processing clinical text data.