A study of entity-linking methods for normalizing Chinese diagnosis and procedure terms to ICD codes

Area of Science:

• Medical Informatics
• Natural Language Processing
• Health Information Management

Background:

• Automated medical coding in China is hindered by variations in physician-documented diagnosis and procedure terms.
• Standardization to International Classification of Diseases (ICD) is crucial for data analysis and healthcare management.

Purpose of the Study:

• To develop and evaluate effective entity-linking methods for normalizing Chinese clinical terms to standard ICD concepts.
• To facilitate automated medical coding in the Chinese healthcare system.

Main Methods:

• Applied an entity-linking framework with two steps: candidate concept generation and ranking.
• Candidate generation used BM25 and an extended version with a synonym knowledgebase.
• Candidate ranking explored BM25, RankSVM, CNN, MatchZoo methods, and a novel BERT-based approach.

Main Results:

• Integrating a synonym knowledgebase improved candidate concept generation coverage to 97.9% (diagnoses) and 93.4% (procedures).
• The BERT-based ranking method achieved the highest accuracy: 92.1% for diagnosis normalization and 80.1% for procedure normalization.
• Performance was evaluated using manually annotated datasets from a Chinese hospital.

Conclusions:

• Developed and compared diverse entity-linking methods for normalizing Chinese clinical terms.
• Demonstrated the feasibility and effectiveness of automated clinical term encoding in Chinese using advanced NLP techniques.
• The BERT-based method shows significant promise for improving the accuracy of automated medical coding.