Stacking-BERT model for Chinese medical procedure entity normalization.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• Medical procedure entity normalization is crucial for semantic medical information sharing.
• Existing methods often rely on context-independent embeddings and lack research in Chinese.
• Challenges include the variety and similarity of medical terms in practice.

Purpose of the Study:

• To develop an effective framework for normalizing Chinese medical procedure terms.
• To improve semantic interoperability in Chinese medical information systems.
• To address the challenges of term variety and similarity in medical entity normalization.

Main Methods:

• A three-step framework: dataset construction, candidate concept generation, and candidate concept ranking.
• Dataset construction utilized external knowledge bases and data augmentation.
• Candidate concept generation employed BM25 retrieval with SNOMED CT synonyms.
• Candidate concept ranking used a stacking-BERT model with adversarial training.

Main Results:

• The stacking-BERT model achieved 93.1% accuracy on the clinical entity normalization task dataset.
• This performance surpassed single BERT models and traditional deep learning approaches.
• Adversarial training and data augmentation proved effective for small-scale training data.

Conclusions:

• The proposed stacking-BERT framework offers an effective solution for Chinese medical procedure entity normalization.
• The study validates the efficacy of BERT-based models in this domain.
• Adversarial training and data augmentation are beneficial for deep learning models with limited data.