Capturing Individual-level Social Determinants from Clinical Text

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Public Health

Background:

• Social determinants of health (SDOH) significantly impact patient outcomes but are often buried in unstructured clinical notes.
• Existing methods for SDOH extraction are limited by narrow definitions and focus on specific note sections.

Purpose of the Study:

• To develop and release a comprehensive dataset for SDOH extraction from entire clinical notes.
• To characterize the documentation of SDOH across various sections of clinical text.
• To evaluate baseline extraction systems and the efficacy of training on limited sections.

Main Methods:

• Creation of a new dataset with 2,007 annotations across 7 SDOH domains within 500 clinical notes.
• Annotation of SDOH across entire notes, not restricted to the social history section.
• Development and evaluation of token classification and generative models for SDOH extraction.

Main Results:

• The study characterizes the presence and location of SDOH information within clinical notes.
• Baseline models demonstrate feasibility for SDOH extraction using both token classification and generative approaches.
• Analysis investigates the effectiveness of training solely on the social history section versus the entire note.

Conclusions:

• A novel, broadly defined SDOH dataset annotated across entire clinical notes has been created.
• The dataset and baseline systems provide a foundation for advancing SDOH extraction research.
• Public release of the dataset will facilitate further investigation into improving patient care through accessible SDOH data.