Enhancing clinical concept extraction with distributional semantics.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Biomedical Informatics

Background:

• Extracting concepts like diagnoses and treatments from clinical narratives is crucial for advanced healthcare applications.
• Supervised learning methods for concept extraction are limited by the scarcity of annotated clinical data.
• Existing methods struggle with the limited scope and patterns covered in labor-intensive annotated training sets.

Purpose of the Study:

• To improve the performance of concept extraction from clinical narratives.
• To address the limitations of supervised learning due to small annotated datasets.
• To integrate empirical semantic relatedness from unannotated text into concept extraction models.

Main Methods:

• Utilized Conditional Random Fields (CRFs), a sequential discriminative classifier, for extracting medical problems, treatments, and tests.
• Employed distributional semantics, analyzing word co-occurrence in a large corpus of clinical trial abstracts (Medline).
• Integrated features derived from word vector representations (cosine metric) and traditional features (dictionary, pattern matching, POS tags).

Main Results:

• The proposed approach significantly improved concept extraction performance compared to a supervised-only baseline.
• Micro-averaged F-score for exact match increased from 80.3% to 82.3%.
• Micro-averaged F-score for inexact match increased from 89.7% to 91.3%, with improvements being statistically significant.

Conclusions:

• Combining supervised learning with distributional semantic features derived from unannotated text substantially enhances concept extraction from clinical narratives.
• This novel approach effectively leverages word distribution information to overcome data limitations in supervised learning.
• The findings demonstrate the significant benefit of distributional semantics for sequence classification tasks in biomedical natural language processing.