Social Determinants of Health Pilot Project in Ohio Uses AI for At-Risk Identification

On Aug. 11 at HIMSS21 in Las Vegas, Lisa M. Lines, Ph.D., senior health services researcher, and Denise Clayton, Ph.D., health economist, from the Research Triangle Park, N.C.-based RTI International, a nonprofit organization, presented an educational session titled “At-Risk Identification Using AI and Social Determinants.”

Lines began the session saying that “The literature and our colleagues all said the same thing: we need better risk-adjustment formulas and performance/quality measures. They don’t take many, if any, social determinants of health (SDOH)—conditions in which people live, work, and grow—into account.”

She also explained local social inequity (LSI), which is a measure explaining health outcome disparities (or inequities) in small geographic areas using predictors related to social factors.

Failing to include SDOH factors can lead to consequences such as practices with lower-risk patients receiving rewards and those with worse-off patients losing out, providers feeling they are penalized for factors outside of their control, and payers and networks having incentives to enroll lower-risk members. Additionally, a lack of good data on SDOH can bias interventions toward lower-risk populations, while current publicly available area-based indices are limited.

Clayton commented that “We need better ways to measure, predict, and adjust for social risk factors in healthcare and population health.”

Lines then introduced the pilot study that took place in Ohio, and later expanded to Kansas, Kentucky, South Carolina, and Tennessee, and the of creation of Artificially Intelligent Risk Adjustment (AIRA) model. The conceptual model shown in the presentation took factors like infant mortality, life expectancy, drug overdose deaths, and excess mortality due to COVID-19 then broke them down into smaller factors like economic stability, healthcare, education, and neighborhood/built environment. Then they were broken down even further into factors like bias, housing, income, transportation and more.

Lines also explained the selected data sources and example measures, breaking them into two categories: integrated application programming interfaces (APIS) and selected downloaded datasets. Some integrated APIs were US DOT transportation measures, FBI’s UCR crime data, and Diversity Data Kids childhood health measures. Some selected downloaded data sets were CDC’s Environmental Public Health Tracking Network, Uniform Crime Reporting Program Data, and the United States Drought Monitor. She then provided a simplified chart of the Random Forest Algorithm, which was used due to generating reasonable predictions across a wide range of data while requiring little configuration.

Clayton presented the findings on applications, splitting four applications into what groups (providers, payers, and policy makers) can use ARIA. Providers, payers, and policy makers can understand drivers of health in order to identify the most important issues to address, use LSI scores to identify individuals or neighborhoods for SDOH interventions, and incorporate LSI scores in evaluations of healthcare innovations, payment models, and interventions on SDOH on higher-risk communities. Payers and policy makers can use LSI scores to risk adjust value-based payment models.

Lines concluded that “Our LSI scores explain 73 percent of the variation in life expectancy in Ohio—an improvement over existing indices that explain 50 to 63 percent. Top individual important factors include child opportunities, receiving food assistance, being raised in two-parent family, property values, and probability of earnings in the top 20 percent (among children born in the same year). She explained further that these measures are complex and multidimensional, covering far more nuance than just “poverty rate.” She said that” We are limited to what data are available, and there may be bias in terms of who is included in the samples used for the underlying measures. While some of the top predictors may track with prior research, others may not be as obvious or amenable to interventions.”

Finally, Lines concluding that “Using information on social risk to explain variation in population health status and outcomes can go beyond just maps.”