Annotating social determinants of health using active learning, and characterizing determinants using neural event...

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Public Health

Background:

• Social determinants of health (SDOH) significantly impact patient health outcomes.
• Understanding SDOH is crucial for effective clinical decision-making.
• Automated extraction of SDOH from clinical text is challenging due to data heterogeneity and the need for detailed annotations.

Purpose of the Study:

• To present a new, detailed corpus for annotating social determinants of health (SDOH) in clinical text.
• To introduce a novel active learning framework for efficient annotation of SDOH events.
• To report initial results of an event extraction model trained on the new corpus.

Main Methods:

• Development of the Social History Annotation Corpus (SHAC) with 4480 social history sections and annotations for 12 SDOH.
• Implementation of an active learning framework using a surrogate text classification task to guide sample selection for annotation.
• Training and evaluation of an event extraction model on the SHAC corpus.

Main Results:

• The SHAC corpus contains detailed annotations for 12 SDOH across 18,000 distinct events.
• The active learning framework effectively increased the identification of health risk factors.
• The event extraction model achieved high performance (0.81-0.93 F1) for substance use, employment, and living status.

Conclusions:

• The SHAC corpus and active learning framework significantly advance the automatic extraction of SDOH from clinical notes.
• Improved SDOH extraction can enhance clinical decision-making and patient care.
• The developed model demonstrates robust performance across multiple institutions.