Using Large Language Models to Identify Social Determinants of Health to Enhance Healthcare Services and Equity
Leveraging AI and natural language processing (NLP) to capture and integrate patient-generated social determinants of health (SDH) data into electronic health records (EHRs) has the potential to enhance patient care by improving the quality of SDH information, enabling actionable clinical decision support (CDS), and reducing the manual burden on clinical staff.
Project Details -
Ongoing
-
Grant NumberR21 HS029991
-
AHRQ Funded Amount$275,000
-
Principal Investigator(s)
-
Organization
-
LocationAuroraColorado
-
Project Dates09/01/2024 - 08/31/2026
-
Care Setting
-
Population
-
Type of Care
-
Health Care Theme
Social determinants of health - the factors that affect the way that people live, grow, work, and age - significantly impact health outcomes and inequalities. The COVID-19 pandemic has further highlighted these disparities, underscoring the need for healthcare systems to address SDH. While capturing this data can enhance patient care and reduce costs, traditional methods like questionnaires often miss the full context of a patient's situation. This limitation has sparked interest in narrative medicine, which allows patients to share personal experiences, offering deeper insights into their social, emotional, and cultural contexts. These narratives can improve care by helping prioritize patient concerns, though challenges persist in integrating them with clinical workflows and evidence. The unstructured nature of narrative data in EHRs presents analytical challenges, but advances in NLP and large language models (LLMs) provide promising solutions for extracting valuable SDH information. These technologies can transform raw narratives into actionable insights, enhancing patient care and enabling more precise interventions. Ongoing initiatives like OurNotes are leveraging patient narratives to improve care. With over 465,000 narratives collected from e-visit check-ins across 200 clinics—covering social history, smoking, alcohol use, and financial strain—OurNotes uses this data to train and evaluate NLP models. This current research, conducted at the University of Colorado Hospital Network in collaboration with Mass General Brigham, is needed for developing and refining tools to better capture and use SDH data in clinical settings.
The specific aims of the research are as follows:
- Develop open ended patient facing questions to capture social determinants challenges.
- Develop LLMs to process relevant SDH information from patient narratives.
- Integrate SDH information into EHR workflows to support CDS.
- Pilot and evaluation of Narrative CDS intervention.
The 5-year research project will begin with a 2-year development phase, followed by a 3-year period focused on single-site EHR integration and pilot testing. Several methods will be employed to achieve the study's aims: developing open-ended patient-facing questions to capture SDH through focus groups with clinicians and patients, integrating narrative data with existing SDH questionnaires during emergency department (ED) self-registration, and creating LLMs to extract and process SDH information from patient narratives. The study will leverage existing annotation schemas and pre-trained LLMs and incorporate structured SDH data into EHR workflows for CDS to trigger interventions based on identified needs. The intervention will be piloted and evaluated in the ED to assess its impact on patient care and SDH documentation quality, with feedback gathered from patients and clinicians on usability, satisfaction, and workflow integration.
The researchers aim to enrich SDH data with valuable context often missing in EHRs, improving the completeness and quality of the information and making it more actionable for patient care. The ultimate goal is to enhance care for at-risk populations while reducing the manual burden on clinical staff.