Unsupervised concept extraction from clinical text through semantic composition.

Area of Science:

• Clinical Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Concept extraction is crucial for enhancing clinical natural language processing (NLP) systems.
• Extracted concepts improve the accuracy and generalization of downstream applications.
• Existing methods often require supervised learning, limiting their applicability.

Purpose of the Study:

• To develop a novel unsupervised system for extracting concepts from clinical text.
• To represent concepts using the Unified Medical Language System (UMLS) and word embeddings.
• To evaluate the system's performance on a standard clinical corpus.

Main Methods:

• The system generates concept representations by integrating UMLS descriptions with word embeddings.
• Higher-order concept vectors are composed from these integrated representations.
• Candidate phrases are identified using a syntactic chunker and labeled with concept vectors.

Main Results:

• The unsupervised system achieved an exact F-score of 0.32 and an inexact F-score of 0.45 on the I2b2-2010 challenge corpus.
• This performance surpasses the only other reported unsupervised concept extraction method.
• The system's reliance on word representations and a chunker ensures it is fully unsupervised.

Conclusions:

• The developed unsupervised system offers a viable alternative for clinical concept extraction.
• Its unsupervised nature allows for application to diverse languages and corpora without prior annotation.
• Open-source code is available, promoting further research and development in clinical NLP.