Clinical Information Extraction From Notes of Veterans With Lymphoid Malignancies: Natural Language Processing Study.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Oncology

Background:

• Clinical natural language processing (cNLP) is vital for extracting data from clinical notes, but its application to rare diseases like lymphoid malignancies is limited by data scarcity and documentation complexity.
• Existing cNLP models may perpetuate biases present in clinical data or development processes, potentially leading to performance disparities.

Purpose of the Study:

• To develop and validate a cNLP pipeline for extracting clinical information from veterans with lymphoid malignancies.
• To assess the pipeline's performance in identifying key clinical entities, including performance status, staging, diagnosis, substance use, and military environmental exposures.

Main Methods:

• A rule-based cNLP pipeline was created, incorporating domain expertise.
• The pipeline's performance was evaluated on clinical notes from two distinct veteran cohorts: non-Hispanic White and non-Hispanic Black veterans.

Main Results:

• The cNLP pipeline demonstrated promising performance, particularly for well-documented entities like performance status.
• Significant racial associations were found in the false-positive and false-negative rates for primary diagnosis detection and in the false-negative rates for substance use identification.

Conclusions:

• The developed cNLP system shows satisfactory and comparable performance for most clinical entities in veterans with lymphoid malignancies.
• Future research should focus on refining the pipeline to address challenges in rare cancer data and mitigate performance biases related to race.