Applying Natural Language Processing to Textual Data From Clinical Data Warehouses: Systematic Review.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Health Data Science

Background:

• Clinical Data Warehouses (CDWs) enable secondary use of health data from clinical care.
• Unstructured clinical text contains significant, high-value information.
• Natural Language Processing (NLP) can structure and improve accessibility of this textual data.

Purpose of the Study:

• To review studies applying NLP to textual data from CDWs.
• To identify common NLP tasks performed on CDW data.
• To categorize the NLP methods employed for these tasks.

Main Methods:

• Systematic review following PRISMA guidelines.
• Searched PubMed, Google Scholar, and ACL Anthology databases.
• Included English articles from 1995-2021 focusing on NLP and CDW textual data.

Main Results:

• 194 out of 1353 articles met inclusion criteria.
• Information extraction (57.7%) and patient identification (26.3%) were primary NLP tasks.
• Symbolic methods (53.4%) were most common, followed by machine learning (30.2%) and deep learning (16.4%).
• Most NLP applications focused on English language data (78.9%).

Conclusions:

• CDWs are crucial for secondary use of clinical text in research.
• Clinical NLP effectively accesses, extracts, and transforms CDW data.
• Challenges persist, especially for non-English languages, but NLP impacts clinical research and practice.