Technology Helps Surface SDOH in Patient Records

By Matt Phillion

What do you do when you have the patient information you need, but that information is buried within the patient record as unstructured data?

This is the challenge NorthShore Edward-Elmhurst Health sought to resolve when they determined they needed a better way to identify patients’ social determinants of health (SDOH) when they presented at the ED—one of the most crucial points in intervention. They knew the details were there in the record, but the challenge was surfacing it when it was most needed.

To overcome this, NorthShore turned to a solution using artificial intelligence and natural language processing (NLP) to delve into the unstructured data in their records to identify patients with SDOH needs when they interacted with the ED and then connect them with appropriate social workers. The social worker then had the opportunity to verify those needs and connect the patient with the necessary community resources or care managements they required.

The big change in process was that this helped ensure social workers would interact with these patients sooner. Previously, they most likely would not have been able to intervene and assist with the patient’s SDOH needs while they were in the ED.

“We had come to understand the significant disparities in our area but we didn’t have a good way of capturing SDOHs,” says Nirav Shah, MD, medical director of quality innovation and clinical practice analytics with NorthShore University Health System.

Leadership was tasked pre-pandemic to look at use cases to help resolve this, and their chief quality officer raised the issue of how to extract SDOH insights from the user notes. They existed in the patient record, but were not in a structured field that could readily impact care.

“The information was buried in the notes,” says Shah. “We knew NLP would be fundamental on our AI journey and from an SDOH standpoint we had social workers, nurses, physicians who were documenting these needs but it was inefficient. There was no way for anyone to know without a deep chart review, which nobody has time to do in addition to the care they’re already providing.”

This led the team to ask: how do we extract the data that’s already in the record?

Modifiable and understandable

The team investigated various solutions and selected Linguamatics, which offered a hybrid platform that was both rules-based and had a machine learning/AI component.

“When we were searching at a time when the AI hype was not at the fever pitch it is now. We liked that this platform was not a black box; it had rules our team could alter and modify,” says Shah. “It was also straightforward enough you didn’t need a data science background. We could democratize it across the system.”

They made use of an existing set of queries and conducted a retrospective validation process, so they didn’t take the results at face value, verifying the data the tool surfaced was accurate and actionable for the patient.

“Knowing we were going to push this out to clinicians who didn’t have a background in data science, we leveraged the same convention used for our scorecards, gold, silver and bronze,” says Shah. “Most of our queries were gold: transportation, food insecurity.”

They chose to go with an on-premise solution rather than in the cloud for security reasons, which helped move the process along.

“All of our notes go into a warehouse and every night we run those notes through this algorithm and push it back into the patient chart. We did this with a six-month lookback for eight or nine social determinants,” says Shah.

They began with ED social workers.

“They are the front line: the ED social workers are trying to close gaps to prevent repeat visits,” says Shah.

They partnered closely with the case management team to design a process that wouldn’t interfere with existing workflows, first using proxies to try to determine SDOH care gaps and later using NLP-enhanced workflows.

There were small adjustments to the workflow. For example, asking the patient if they have a specific issue and entering this into the flowsheet rows that went back to the data scientists. All this validation helped ensure the algorithm was turning up true results.

“One example: we had a depression query that included identifying weight gain, which can be a sign of depression,” says Shah. “But every pregnant patient was getting flagged—and while depression and pregnancy may be an issue, this was flagging every single patient, so we were quickly able to modify the query,” says Shah.

As a result of this instance, the ED social workers came to trust in the system.

“We used to have to comb through the notes. If it’s an algorithm, it will identify if they have a care gap: does this patient have an unmet need? And you can click right to where it is in the note,” says Shah. “Where they were spending 80% of their time on chart review, they now can spend that 80% making an impact on the patient.”

They asked staff using the solution for cases where they felt it had an impact, and Shah notes that surprisingly, the examples the staff gave were all victims of crime or abuse.

“You’d think it would be a broad array of things, but these were factors they were not able to pick up in advance. It was novel identifying an issue that might be related to symptoms,” says Shah. “A 20- or 30-year-old person might not trigger as being high risk for mortality but they got flagged for high risk for abuse and crime. There were non-specific complaints that, once collected, changed how they approached the patient and they were able to connect them with support and services when it might otherwise have gone unnoticed.”

Shah notes that new innovations the organization introduces are meant to help everyone practice at the top of their license, but staff have called this solution a game changer.

“The impact was immediate and palpable. You can imagine how inefficient the process of reviewing the chart was before,” says Shah.

This is particularly impactful with older patients who may not have immediately apparent SDOH issues but also have an extensive chart to review.

“Now it’s right in your face: they have an issue, you are taken to where it is in the text, and you can go make an impact,” says Shah.

Continued improvement to ensure accuracy

There’s been some adjustments to the algorithm to help with accuracy and effectiveness. They continue to meet with their users and work with data scientists to ensure the validation process is working and results continue to be more efficient and accurate.

“There’s some work up front. They fill out a flow sheet for the patient at the start, but it takes a ton of work out later on with chart review,” says Shah.

They also benchmarked results against U.S. census data to ensure accuracy, and have found for most categories—food insecurity, transportation, housing—the results are very close.

“That validates that we’re documenting things appropriately and the tool is picking it up,” says Shah.

Knowing the tool is working, the future looks bright. The team is spending a lot of time socializing the technology and educating staff about its capabilities across the organization.

“That takes time,” says Shah. “Diffusion is just as important for growth as innovation. In healthcare we do a lot of ‘new, shiny object’ stuff but we don’t spread it around once the pilot loses energy, so we need to diffuse this across our system. We have this SDOH insight, so we’re now in talks with navigation, primary care, inpatient medicine. Eventually everyone will be required to collect this information, so how do we leverage this workflow?”

Conversation with other leaders has led to interest in getting it into the hands of clinicians and even research teams.

“Just having access to SDOH insights on every patient in a structured way is like having a career’s worth of research,” says Shah. “How do we get that into the right hands?”

There’s also potential for the technology in other areas like registry automation for oncology registries, as well as mental health.

“Chart review in that area is a lot of manual labor,” says Shah.

The potential isn’t limited to any one area, Shah says.

“We’re building a governance structure around this so we can disseminate it out and crowdsource ideas,” he says.

Matt Phillion is a freelance writer covering healthcare, cybersecurity, and more. He can be reached at matthew.phillion@gmail.com.