Normalizing clinical terms using learned edit distance patterns.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Clinical text variations pose challenges for standard terminology matching.
• Existing normalization methods lack accuracy and comprehensive coverage.
• Automated pattern learning offers a potential solution for clinical term variation.

Purpose of the Study:

• To develop a novel method for automatically learning clinical term variation patterns.
• To normalize previously unseen clinical terms using generalized patterns.
• To evaluate the method's performance on disease and disorder mention normalization.

Main Methods:

• Learned variation patterns by computing edit distances on known variations from the Unified Medical Language System (UMLS).
• Generalized learned patterns to normalize unseen clinical terms.
• Evaluated performance on the SemEval 2014 dataset, comparing against MetaMap and cosine similarity.

Main Results:

• Achieved 64.7% accuracy in normalizing disease mentions not exactly matching UMLS or training data.
• Outperformed MetaMap (41.9% accuracy) and cosine similarity (44.6% accuracy).
• Obtained a 54.4% best F-measure for CUIs, significantly higher than MetaMap (19.4%) and cosine similarity (38.1%).

Conclusions:

• The novel method demonstrates superior performance in normalizing clinical disease mentions compared to MetaMap and cosine similarity.
• The approach effectively handles terms not present in standard terminologies like UMLS.
• The method is generalizable for normalizing clinical terms across various semantic types.