How an ‘Artificially Intelligent’ Local Social Risk Score can Improve Population Health on the Path Toward Health Equity

This is the second part of a two-part blog series on addressing social determinants of health.

Robust population health management requires identifying populations at high risk for unmet social needs. While COVID-19 has highlighted the widespread inequities in our current health system, data analytics on populations who may not regularly access health care will enable targeted initiatives to both improve health outcomes and inform payment models. 

There is a real need for scalable partnerships between health care providers and community-based organizations to meet the identified social needs of individuals receiving medical care. But in our fragmented health care system that exists today, it is not enough. People who lack access to medical care, health literacy, and/or trust in the health system may not be screened for social risks until it is too late. In part due to health inequities, this same population is more likely to have chronic health conditions, lack a primary care provider, and make potentially avoidable visits to emergency departments. As mentioned in the first part of this blog series, it may be challenging to identify and gain the trust of this population for COVID-19 vaccine rollout. 

Neighborhood-Level Measurement of SDoH Needs Complements Individual Data

Fortunately, population health management does not have to rely only on person-level information. Neighborhood-level measurement of SDoH needs can complement individual data. Ideally, individual needs and neighborhood characteristics are both available. Yet individual information is often missing - whether because people are not accessing care, or because physician offices do not have adequate time, systems, and staffing to screen for SDoH needs, or because of privacy concerns. In these cases, neighborhood characteristics can guide efforts to reach and care for higher-risk populations. Importantly, neighborhood characteristics may sometimes proxy individual characteristics, but are often independently predictive of poor outcomes.

Place-based indices of social risk are not new, but those commonly used include only a small number of variables (fewer than 20) and have limited predictive power. Motivated by the need to incorporate more granular data into risk predictions, RTI has developed an “artificially intelligent” approach to assessing which neighborhood-level social drivers of health are most predictive of neighborhood-level life expectancy. 

We used a machine-learning method known as random forests (RFs), which have several big advantages over traditional regression-based approaches. One is that you can include hundreds of variables. Our Local Social Risk score was piloted using over 140 publicly available variables for Ohio. The 5 most important predictors were:

1. percentage of the population that received food assistance 
2. median owner-occupied property values
3. probability of reaching the top quintile of the national household income distribution (among children born in the same year)
4. having a college degree
5. percent enrolled in Medicaid

Local Social Risk Score Has Many Important Applications

Our Local Social Risk score explains 73% of neighborhood variation in Ohio’s life expectancy, where the range is 60 to 89.2 years. This is a substantial improvement over alternative, commonly used indices (such as the Area Deprivation Index, Social Deprivation Index, and Social Vulnerability Index, which explain only 50-63% of the variation in life expectancy in Ohio).  

The Local Social Risk score has many potential applications, including:

• Identifying neighborhoods at highest risk of poor outcomes for better targeting of interventions and resources 
• Understanding the impact of healthcare innovations, payment models, and interventions on social drivers of health in high-risk communities
• Accounting for factors outside of providers’ control for more fair and equitable performance/quality measurement, evaluation, risk adjustment, and reimbursement

With the recent spotlight on inequities in COVID-19 outcomes and the resulting economic downturn, better data on social drivers and risks are urgently needed. Both community-level and individual-level approaches are called for if we want to see systemic change in addressing health-related social needs and improving health equity. Beyond COVID-19, if we want to improve population health and reduce risk in a value-based health care system, we must account for SDoH. Using better data and making better predictions can help identify where to concentrate efforts, taking us one step further toward this goal.