A study of abbreviations in clinical notes

Area of Science:

• Natural Language Processing (NLP)
• Clinical Informatics
• Biomedical Data Mining

Background:

• Natural Language Processing (NLP) systems aim to extract patient information from clinical notes for applications like error detection and decision support.
• Clinical notes frequently use undefined abbreviations, posing a challenge for NLP interpretation compared to biomedical literature.
• Accurate interpretation of abbreviations is critical for leveraging clinical data.

Purpose of the Study:

• To develop and evaluate a two-step model for building a clinical abbreviation database.
• To identify effective methods for detecting abbreviations in clinical text.
• To create and assess sense inventories for identified abbreviations using knowledge sources like UMLS and ADAM.

Main Methods:

• Developed and evaluated four distinct methods for detecting abbreviations within a clinical text corpus.
• Implemented a method to build abbreviation sense inventories using the Unified Medical Language System (UMLS) and a MEDLINE abbreviation database (ADAM).
• Evaluated the performance of abbreviation detection and the quality of the generated sense inventories.

Main Results:

• The best abbreviation detection method achieved 91.4% precision and 80.3% recall.
• The Unified Medical Language System (UMLS) provided a more suitable sense inventory than ADAM, covering 35% of senses with a 40% ambiguity rate.
• A significant portion of abbreviations and their correct senses remain uncovered by automated methods.

Conclusions:

• A two-step model shows promise for building clinical abbreviation databases, with effective abbreviation detection methods identified.
• While UMLS is a valuable resource, it has limitations in coverage and ambiguity for clinical abbreviations.
• Domain expert annotation remains essential for comprehensive and accurate interpretation of clinical abbreviations in NLP systems.