Data-Driven Methods to Identify Social Determinants of Health

Background: There is increased attention on social determinants of health (SDOH) as a result of empirical evidence showing that the patient?s social background is associated with their health behaviors and clinical outcomes. Now more than ever, health care systems (HCS) are being held accountable for addressing social factors.

Improving the quality of health care among racial and ethnic minorities is a VA is a top priority.

Significance/Impact: Ideally, identifying and documenting a patient?s social background would be followed by referral to services that address the SDOH that are most likely to reduce compliance with recommendations for disease prevention, treatment, and management.

However, SDOH such as education, income, social isolation, and financial strain are rarely documented during routine care visits.

A more systematic approach that leverages health information technology is needed to improve the efficiency and effectiveness of identifying social determinants among patients in the VA so that more targeted approaches are used to address these risk factors in the patients? communities.

A better understanding of SDOH within the electronic health record (EHR) is needed in order to improve population health management and processes for referring patients to social services.

Innovation: The first step to developing a more robust data-driven strategy for identifying social phenotypes among patients is to understand the extent to which SDOHs are being documented in the EHR.

Natural language processing (NLP) is one strategy to automatically extract those data from clinical notes in the EHR into a structured format that can be used to examine the quality of health care and facilitate the development and implementation of quality improvement strategies.

However, NLP approaches alone are not sufficient to improve the quality of health care for Veteran racial/ethnic minorities.

This is because poor quality communication between patients and providers and greater distrust in the health care system among minorities may limit discussion of these factors.

Novel deep learning approaches have not been fully leveraged in the identification of patients at risk for adverse SDOH.

Moreover, there is a lack of empirical data on the concordance between patient self-reported SDOH and those extracted using NLP. Even less is known about the value associated with obtaining and documenting SDOH on patient outcomes.

Therefore, we propose to develop a multilevel health informatics approach for identifying social phenotypes among primary care patients based on documentation of SDOH in the EHR as part of the following: Specific Aims: Aim 1: Use deep learning strategies to identify social phenotypes among diabetes patients based on documentation of SDOH in the EHR.

Aim 2: Examine the concordance between risk factors for SDOH identified using NLP and patient-self- report. Aim 3: Conduct a study to evaluate the effects of documenting SDOH on patient outcomes.

Methodology: A deep learning NLP approach will be used to characterize the rates at which SDOH are documented in the EHR. Machine learning strategies will be used to identify social phenotypes based on SDOH.

Implementation/Next Steps: We predict that Veterans who have SDOH documented in the EHR will report better clinical outcomes, greater trust in health care providers, and better patient-physician communication compared to Veterans who do not have SDOH documented in their EHR. We will also characterize referrals to clinic- and community-based services based on the patient?s social phenotype.