DrAfter123/DigitalVision Vectors

How predictive analytics could improve SDOH screening

As healthcare organizations seek to improve health equity, predictive analytics using robust EHR data could help flag patients likely to have future health-related social needs.

Predictive analytics is used for a variety of high-value use cases in healthcare, from clinical decision support to population health management, as it allows healthcare organizations to predict what is likely to happen in the future based on historical data.

Now, researchers are exploring how machine learning (ML) could help healthcare organizations predict and proactively address patients' health-related social needs, also known as social determinants of health (SDOH).

Mitigating SDOH is critical in the pursuit of health equity, with social factors like food insecurity and housing instability accounting for between 30% to 55% of health outcomes. However, clinical documentation burden and poor workflow integration make standardized SDOH screening challenging for many healthcare organizations.

Tapping EHR data for SDOH prediction

With growing amounts of patient information available within the EHR, could health systems use existing data and machine learning to predict the need for SDOH interventions?

"I've always been interested in, 'How do we take information and squeeze the most value out of it?'" said Joshua Vest, Ph.D., M.P.H., a Regenstrief Institute research scientist and professor of health policy and management at the Indiana University Indianapolis Richard M. Fairbanks School of Public Health. "'How do we take interesting information and do interesting things with it to improve health and population health and public health?'"

Vest noted that health systems already collect a lot of information about individuals and much of this data has implications for health-related social needs.

For instance, changes in patient insurance could indicate changes in employment status, and a constantly changing residential history could suggest housing instability.

I've always been interested in, 'How do we take information and squeeze the most value out of it? How do we take interesting information and do interesting things with it to improve health and population health and public health?'
Joshua Vest, Ph.D., M.P.H.Research scientist, Regenstrief Institute

"Classically, you could look in a historical EHR record and see, 'How many times is somebody moving around, or are their addresses really aliases for homelessness? Is their address a known shelter in your community, or is it a known church, or is it even the address of the ER hospital itself?'" he explained.

Other indications of health-related social needs include referrals to social services like SNAP and WIC that suggest food insecurity.

Anticipating SDOH with ML

In a new PLOS One study, Vest and a group of researchers applied ML predictive modeling to existing EHR data sources to predict patients' likely need for health-related social services in the next 30 days.

The researchers trained the ML predictive model using EHR data from 1,101 adult patients who sought care at the Eskenazi Health emergency department (ED) in Indianapolis, between June 2021 and March 2023 and completed an SDOH survey.

The model applied previously validated natural language processing (NLP) algorithms to analyze clinical data and identify indicators of health-related social needs. Additionally, the ML model incorporated area deprivation index scores based on patient addresses and linked patient data to a large multi-institutional clinical repository to gain clinical insights from outside Eskenazi Health.

Then, the researchers compared the performance of the predictive ML model with the performance of a classification algorithm using SDOH screening responses.

Overall, the ML model outperformed the screening questionnaire model in predicting patients' need for health-related social services, signaling the potential of predictive analytics to help healthcare organizations more proactively anticipate patients' SDOH needs.

Addressing bias in SDOH prediction models

Notably, both models exhibited bias, performing higher in predicting SDOH needs for white, non-Hispanic groups. According to Vest, this bias is largely a function of data and unequal access to care.

"Not everybody has the same access to healthcare in the United States -- there's variance by your insurance, variance by where you live, variance by linguistic challenges," said Vest. "Just purely relying on secondary data and models, we're probably capturing the fact that some people have more data available to them than others simply because they're able to get to care, so that's a challenge that's a little hard to overcome."

However, Vest noted that statistical methods can help correct data variances that lead to ML bias.

"We lean a lot on the great work that's being done in computer science and informatics, how they're thinking about prediction, and how they're making sure that their models are fair," said Vest. "We try to take those models and adapt them so when we roll stuff out actually to try out or in trials, we correct to make sure that nobody is being disadvantaged by performance."

Vest noted that while the ML model trained on EHR data performed better than the screening questionnaire model, adding age and gender to the screening questionnaire would have improved the model's performance.

"We think that's a pretty interesting concept that, for example, individuals' financial needs vary and their perceptions of their financial needs vary dramatically over the course of life," he said. "Somebody who's in their early twenties has a very different financial situation and constraints than somebody who is 65 and older and retired, so this idea of asking the same questions to everybody doesn't really capture that potential variation that exists in individuals' underlying risk."

Additionally, he pointed out that women are more likely to experience housing instability than men due to violence issues and employment challenges.

"Thinking about how we can start incorporating some of the additional information into the screening questions would be a way to improve them," Vest said. "For instance, maybe incorporating age into the screenings or just considering that people may have different levels of risk."

Future research directions

The research team is preparing to launch a trial of the ML model next month at a clinical partner's ED. The trial will integrate the model into a standardized platform that consolidates data from a health information exchange within the EHR.

"We have been building our risk prediction models to fit into that existing platform, so it's a single sign-on integrated access to additional information around social needs that goes along with all the health information exchange," Vest said. "Hopefully, we'll be rolling that out next month to see if it has effects on referral rates, screening rates and subsequent utilization."

Moving forward, the research team is also evaluating how predictive modeling compares to NLP approaches in identifying SDOH needs.

"We're really trying to do the underlying psychometrics of these tools and all these options for measuring health-related social needs and see how they really perform and if there are important variations across populations and settings," Vest said.

Hannah Nelson has been covering news related to health information technology and health data interoperability since 2020.

Dig Deeper on Predictive analytics in healthcare