Integrating Social Determinants of Health Into Risk Prediction Models

While many risk factors patients face concerning their health are due to genetics, researchers suggest that social determinants of health could also play a role in risk prediction.

In a recent study, researchers examined how different social determinants of health (SDoH) impact risk prediction models.

“We’ve been interested in risk prediction for some time. My particular interest is in helping identify patients who have social risks and getting to the services they need to address those risks. Risk stratification of modeling is a key first step in that process,” study co-author Joshua Vest, PhD, of the Regenstrief Institute and Indiana University Richard M. Fairbanks School of Public Health told HealthITAnalytics.

The research team set out to provide empirical guidance by comparing the performance of six different area-level social determinants of health measurement approaches in predicting patient referral to a social worker and hospital admission following a primary care visit.

“Social determinants are such an important part of healthcare now, and a big focus of addressing population health issues and responding to policy imperatives. More and more researchers, healthcare organizations, and government agencies are trying to leverage social determinants data in a variety of applications,” Vest said.

According to the study, informatics and clinical researchers are increasingly shifting their focus to area-based social determinants of health to analyze patient-level clinical datasets. While SDoH are not always the most reliable data points to measure an individual’s characteristics, information regarding their living environment can provide significant insight into a patient’s health.

“When you think about what drives health, what defines whether we are healthy and have a high quality of life, it’s not medical care. Medical care is important, but medical care occurs after you’ve gotten sick. Medical care is reactive. Medical care is only a small proportion. What is much more important are your individual behaviors, the environment you live in, the choices you make, your social support system, the resources you have available to you,” Vest explained.

Using a random forest classification algorithm, the researchers examined different SDOH measurement approaches in predicting patient referral to social workers and hospital admission after a primary care visit.

The research team used a sample consisting of 209,605 adult patients from Eskenazi Health’s Federally Qualified Health Center (FQHC) primary care clinic from 2016- 2018. The FQHC’s nine sites serve the Indianapolis metropolitan area.

Additionally, the researchers utilized four data sources. Individual-level healthcare encounters, demographics, and clinical indicators were obtained from the FQHC’s electronic health records (EHR) from the Indiana Network for Patient Care (INPC). The researchers receive the area-based SDOH measures from the US Census Bureau’s American Community Survey (ACS).

The team then linked the ACS five-year estimated to patient records based on residential census tract. Lastly, the Indiana State Department of Health provided area-level mortality data.

“What we sought to do was try to figure out an effective and efficient way to use that data. Particularly, you can measure social factors at an individual level. You can also try to understand the patient’s context and understand by using aggregate measures. Those aggregate measures are common, easy to get, and can be informative,” Vest said.

After analyzing the performances of a series of random forest prediction models using the patient-level features in combination with the different area-level feature sets, researchers found that 5.4 percent of primary care encounters resulted in a referral to a social worker in the next year and 1.8 percent resulted in hospital admission within seven days.

“Adding area-level measures of SDoH to patient-level information improved the performance of random forest models predicting patient need for a referral to a social worker. Improvement in models predicting hospital admissions was limited to higher specificity. Model performance was consistent with, or potentially better than, other prediction models focused on population health relevant measures and utilization outcomes,” the authors wrote in the study.

According to Vest, the biggest takeaway from the study was that the best approach to conducting research does not have to be the most complex approach. Doing so makes information and data more accessible to other organizations and researchers.

“We’ve tried multiple different ways to look at this data, to understand, to create more informative approaches. What we found was just using the most basic approach of the individual features worked out very well. More complex, computationally intensive or other theoretical-driven approaches aren’t necessarily required. It just makes the data a lot more accessible to individuals for their needs,” Vest said.

To develop effective and accurate methods of predicting risk in individuals, researchers need to train machine learning technology to consider social determinants of health.

“Social determinants of health are influential on health and wellbeing. If we can support patients and individuals and populations through that, we can avoid poor health, we can avoid unnecessary utilization, we can avoid costs,” Vest said.