Extracting similar terms from multiple EMR-based semantic embeddings to support chart reviews

Area of Science:

• Natural Language Processing
• Medical Informatics

Background:

• Word embeddings represent semantically similar terms in a vector space.
• Clinical text embeddings can enhance keyword search and text highlighting in electronic medical records (EMR).

Purpose of the Study:

• To present methods for refining similar term selection from multiple EMR-based word embeddings.
• To quantitatively and qualitatively evaluate the performance of refined terms across chart review tasks.

Main Methods:

• Trained word embeddings on individual clinical note types within an EMR.
• Combined, weighted, and truncated embeddings to select refined similar terms.
• Evaluated term quality using information retrieval metrics (precision-at-K) and user studies.

Main Results:

• Refined terms demonstrated superior information retrieval performance compared to baseline methods (e.g., average P@5 increased from 0.48 to 0.60).
• User studies indicated a preference for refined terms.
• A timing study showed reduced time to answer questions using refined terms.

Conclusions:

• Semantically similar terms derived from multiple word embeddings accelerate clinical information retrieval and synthesis.
• The refined term selection process offers a significant improvement for clinical text analysis and chart review.