Toward Granular Social Determinants of Health (SDoH) Coding: A Semantics AI Framework to Extract and Encode SDoH enabled by Large Language Models

Poster Number: P95
Presentation Time: 05:00 PM - 06:30 PM

Abstract Keywords: Large Language Models (LLMs), Natural Language Processing, Information Extraction, Controlled Terminologies, Ontologies, and Vocabularies
Primary Track: Applications
Programmatic Theme: Clinical Research Informatics

We propose a framework that leverages Social Determinants of Health (SDoH) ontological knowledge to guide a large language model (specifically GPT-4) to identify SDoH elements from free-text electronic health record notes and map them to standard granular SDoH concepts. Here, we leverage Intelligent Medical Objects (IMO) coding of granular SDoH content collected through real world patient problem list. First, given a note and broad SDoH categories, we prompt GPT-4 to identify all SDoH elements and classify each element to a broad category along with providing a rationale and the evidence note text. Subsequently, we prompt the model to select final granular concepts given its previous response. Our framework achieved an F1-score of 90.18 in identifying SDoH elements, when evaluated on 30 MIMIC-III notes. Among the correctly identified elements, the accuracies in classifying the broad SDoH category and mapping to granular IMO concepts were 86.63% and 95.45%, respectively.

Speaker(s):
Surabhi Datta, PhD
Intelligent Medical Objects

Author(s):
Surabhi Datta, PhD - Intelligent Medical Objects; Hunki Paek, PhD - Intelligent Medical Objects; Kyeryoung Lee, PhD - Intelligent Medical Objects; Liang-Chin Huang, PhD - Intelligent Medical Objects; Jingqi Wang, PhD - Intelligent Medical Objects; Xiaoyan Wang, PhD - Intelligent Medical Objects;