A Hybrid Normalization Method for Medical Concepts in Clinical Narrative using Semantic Matching

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Clinical term normalization is crucial for standardizing medical notes.
• Existing methods struggle with semantic similarity and concept mapping.
• A significant portion of clinical mentions lack direct concept mapping in current systems.

Purpose of the Study:

• To develop a hybrid normalization system integrating deep learning with dictionary lookup.
• To enhance the capture of semantic similarity between clinical concept expressions.
• To improve the accuracy of mapping clinical terms to standardized medical vocabularies.

Main Methods:

• Developed a hybrid system combining deep learning models with dictionary lookup.
• Incorporated semantic similarity analysis to complement traditional approaches.
• Evaluated the system on the ShARe/CLEF 2013 challenge dataset.

Main Results:

• Achieved 90.6% accuracy in normalizing mentions to existing concepts.
• Demonstrated a statistically significant improvement of 2.6% over baseline methods.
• Identified inconsistencies in challenge data and ambiguities in the Unified Medical Language System (UMLS).

Conclusions:

• The hybrid deep learning approach significantly enhances clinical term normalization accuracy.
• Semantic similarity analysis is key to improving normalization performance.
• Further research can leverage deep learning for more robust medical term mapping.