Expanding a radiology lexicon using contextual patterns in radiology reports.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Medical Informatics

Background:

• Clinical notes contain valuable information but suffer from inconsistent terminology.
• Biomedical lexicons are crucial for information retrieval but are costly to create and maintain.
• Distributional semantics offers a potential solution for efficient lexicon expansion.

Purpose of the Study:

• To investigate the efficacy of distributional semantics in accelerating lexicon expansion within the clinical domain of radiology.
• To identify synonyms for terms in the RadLex lexicon using word2vec applied to radiology notes.

Main Methods:

• Utilized the word2vec software package to process a corpus of radiology notes.
• Applied distributional semantics to identify word and phrase representations.
• Stratified performance analysis based on term category, frequency, length, vector magnitude, and context window.

Main Results:

• Over 50% of synonyms were found within the top 25 ranked terms from a list of over 775,000.
• Synonym identification was more effective for multi-word terms and terms with higher corpus frequency (>=100).
• Term vector magnitude between 4 and 5, and certain RadLex categories (e.g., anatomical substances) showed better synonym discovery.

Conclusions:

• Distributional semantics algorithms can significantly enhance the efficiency of clinical lexicon curation.
• This approach offers a cost-effective and time-saving alternative to manual lexicon development.
• Leveraging unstructured clinical text with NLP techniques improves access to medical information.