How AI and machine learning can help predict SDOH needs

In his upcoming HIMSS22 session, Dr. Jim Walton, president and CEO of Genesis Physicians Group, will give an in-depth look at his work with predictive machine learning and social determinants of health.
By Bill Siwicki
11:18 AM

Dr. Jim Walton, president and CEO of Genesis Physicians Group

Photo: Genesis Physicians Group

Social determinants of health are major contributors to health inequity and rising healthcare costs in vulnerable populations such as Medicaid beneficiaries.

Healthcare innovators are building proactive care management programs to mitigate SDOH risk by connecting high-risk members with community-based organizations to arrange food delivery, transportation to appointments, emergency housing and other services.

With limited care-management resources available, organizations are turning to artificial intelligence to accurately identify high-risk members with addressable SDOH and efficiently target interventions.

In his March HIMSS22 session entitled " Using Explainable AI to Mitigate SDOH Contributors to Risk," Dr. Jim Walton, president and CEO of Genesis Physicians Group, will describe how his organization, along with Medical Home Network, avoided potential pitfalls of applying AI in underrepresented populations and trained machine learning models on the population and data sources to fairly and efficiently identify high-risk members.

Healthcare IT News interviewed Dr. Walton to get an advance look at his upcoming educational session.

Q. How do care-management interventions to mitigate SDOH as a risk for undesired health outcomes work?

A. Recently, population-health-management strategies have begun to incorporate evaluations for patients' social needs connected to SDOH, as well as interventions addressing these needs. These interventions rely on the organizational capability to accurately assess individual social needs and provide timely responses to these needs in order to improve clinical outcomes.

These new strategies have emerged over the last few years and have accelerated as the COVID-19 pandemic has highlighted the reality of disparities in health and health outcomes among minority populations. In short, new organization- and provider-level emphasis on including SDOH along with traditional clinical diagnosis and utilization data is helping to "round out" the picture of patient populations targeted for care-management interventions.

Care managers and social workers, working within accountable care organizations, and physician provider networks now incorporate a short series of SDOH interview questions with patients identified as high-risk for poor healthcare outcomes or unnecessary future healthcare expenditures.

The SDOH questions are designed to surface pressing social needs that patients are experiencing that might be impinging on their overall health status or prohibiting them from accomplishing their health goals. The care-management team members work to connect patients and their families to community-based organizations that offer solutions for many of the social needs identified.

In turn, the patients begin to see the organization and/or physician network as more trustworthy and credible, as whole-person care is substituted for the chronic disease care model of the past.

Historically, care-management services have focused on clinical disease management, patient education, appointment navigation and pharmaceutical adherence issues that chronically ill patients face each day. 

Now, we see the evolution of care-management services that are more agile, where interventional staff are just as likely to identify and respond to the social needs that many patients express as they are the clinical disease management.

It is as if the expressed social need is now becoming recognized as the real barrier to realizing health goals – for example, completing a preventive service like breast, colorectal, prostate, cervical cancer screening or successfully controlling a chronic disease condition through medication adherence.

By reducing or eliminating expressed social needs, care managers gain more credibility in the minds of the patients, who then are more likely to hear and/or adhere to recommendations on how to maintain their health, rather than only responding when there is an acute health crisis or exacerbation of a chronic disease.

Social needs may include such diverse issues as finding a trusted childcare resource to give a parent time to keep appointments, affordable transportation, food stability, housing stability, et cetera. These social needs have not always been viewed as the purview of the traditional healthcare delivery system.

That said, hospital social workers and nurse discharge planners are accustomed to having to help patients with these types of issues during an inpatient admission. Now, we see these social interventions occurring as a matter of daily work for accountable care organizations and physician networks participating in value-based payment arrangements with both commercial and government payers because quality, cost and patient satisfaction measures are key elements of their contract and connected to financial rewards.

Q. Please describe the process for using local data to train predictive machine learning models.

A. A specific population's social needs can be identified using publicly available social data related to a person's address – census tract and/or ZIP code level data. While these data are directionally accurate, they are not an adequate substitute for individual patient-reported data around expressed social needs that often create barriers to access to healthcare as well as the outcomes of healthcare.

Additionally, a patient's clinical data – for example, utilization, costs, pharmaceutical utilization – can be mixed with both population-based social determinant data and individually reported social need data to create a more complete risk profile stratification process for a specific population.

With machine learning technology, data scientists can risk-stratify the population, placing patients with higher burdens of social risk impacting their health access and outcomes at the top, and those with less burden toward the bottom.

As organizations and provider networks intervene on identified social needs, the resulting changes in both clinical outcomes and social needs can be used in a feedback loop to retrain the machine learning algorithm, helping the model become more precise in determining which social need intervention may have the highest likelihood of producing the greatest positive impact, improving efficiency for the intervening organization.

Q. What are a couple of approaches to AI that maximize predictive accuracy?

A. One of the most important ways to maximize predictive accuracy is to train custom AI/machine learning models on a specific population with available data sources, instead of using an off-the-shelf model trained on a general population.

There are three main reasons for this. First, the accuracy of any predictive model drops when the model was trained on a population that is not representative of the population in which predictions are made. For example, training a predictive model on the general population may be inaccurate when used in a Medicare or Medicaid population.

Second, off-the-shelf models often are trained on certain data types or data sources, so if your organization doesn't have access to the same type of data, predictive accuracy may drop. 

The third reason is the converse – your organization may have access to additional data types that are not included in the typical off-the-shelf model, which means that the model is not taking full advantage of data that may potentially contribute to better predictive accuracy.

As an example, Genesis Physicians Group conducts individual interviews or surveys around SDOH and social needs that are highly connected to the risk of future adverse events that aren't easily incorporated into off-the-shelf predictive models.

There are some instances where a custom model might not offer improved accuracy. If a population is too small and/or the outcome we are interested in predicting is very rare, we may not have enough events of the outcome of interest to sufficiently train a custom model.

Dr. Walton's HIMSS22 session, "Using Explainable AI to Mitigate SDOH Contributors to Risk," is scheduled for March 15 from 1:30-2:30 p.m. in the Orange County Convention Center in room W303A. His co-presenter is Cheryl Lulias, president and CEO of Medical Home Network.

Twitter: @SiwickiHealthIT
Email the writer: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.

Want to get more stories like this one? Get daily news updates from Healthcare IT News.
Your subscription has been saved.
Something went wrong. Please try again.