NorthShore – Edward-Elmhurst Health has been developing a natural language processing-based system to extract social determinants of health (SDOH) information from unstructured data in the EHR for use by social workers in the emergency room.
Nirav Shah, M.D., M.P.H., medical director of quality innovation and clinical practice analytics for the Chicago-based health system, recently spoke with Healthcare Innovation about their plans for using this innovation.
In January 2022, NorthShore University HealthSystem and Edward-Elmhurst Health completed their merger. The combined health system has nine hospitals and more than 300 ambulatory locations across six counties in northeast Illinois.
Although the Epic EHR has structured fields where clinical teams can enter SDOH information, few clinicians are using it so far. The NLP effort started with Shah’s team getting an internal grant and being tasked by health system leadership to extract social determinants of health information from the EHR's unstructured notes. They spent time looking at the vendor space to see if they could build something to extract that data. “We ended up partnering with a company called Linguamatics, which was purchased by IQVIA. What's interesting about their tool is that it's rules-based,” Shah explained. “We were also looking at deep learning tools, but at the time when we were searching, a lot of those focused on specific use cases like social determinants, but then expanding to other things was challenging. This is very customizable.”
Previously, they used to identify high-risk patients with a model that looked at high risk for readmission and high risk for mortality. Or if a nurse identified some type of SDOH issue that should be addressed by the social worker, they would flag it.
Now, they are getting additional information that includes a look-back from any clinical note in the system from the past six months. It could be from a primary care physician, caseworker, social worker, or nurse. “There were specific domains that we looked at, including housing insecurity, food insecurity, transportation issues, domestic violence, and stress,” Shah said.
With this data linked back into Epic, the ED social workers helped validate in the notes whether something was accurate or not, which helps fine-tune the queries. “To give you an example, under depression and mental health, one of the sub-queries to that was weight gain, because it's often part of the syndrome related to depression. But it was flagging pregnant women for gaining weight,” Shah said. “And as a result of that, because we were able to pick it up immediately, we fine-tuned the queries to eliminate the piece of the query that was doing that. I think this is extremely critical, because it's not a black box. We figured out what was going on, and there was immediate feedback and we changed it. This type of prospective validation and the connection with our data scientists builds that trust, so the ED social workers now really believe in the process.”
The NLP tool picked up a bunch of domestic violence, but there are a lot of sub-queries, so it's not all necessarily domestic violence, he added. “One of the things that it picked up was potential neglect in a skilled nursing facility, so we were able to change the patient's nursing facility as a result of that.”
The SDOH concerns the clinical notes are picking up are pretty closely matching up with U.S. Census data. For instance, the Census has 8.5 percent of people reporting a transportation issue; NorthShore picked up 9.3 percent. But they also got qualitative feedback from social workers. “One social worker said this is a game changer because literally everything that is getting flagged by this NLP process is intervenable, whereas that was not the case with the prior workflow because we were using a proxy that's not a one-to-one match with an SDOH gap,” Shah said.
Shah said this was just their first use case for the tool. “We're actively working with our primary care colleagues, our navigators and our OB colleagues; these are areas where we think we can create broad impact, trying to figure out how to deploy this,” he added. “Can we leverage this to identify patients who need a deeper kind of survey around these issues that would then fit within the Epic survey questions that they have?”
NorthShore – Edward-Elmhurst Health also is thinking about using this at the population health level. “Can we run analytics on this for the entire population in our catchment area to identify where hot spots are, where we may need to address transportation issues, or there are issues around housing instability or food insecurity?” Where can we place additional social workers in clinics where there may be more need as opposed to just determining based on ED visits or something a proxy that may or may not get us to exactly what we need?”
They also are trying to figure out to add more automation around registries such as oncology registries. “We are looking at how we could speed up this type of workflow, which is fairly manual. How can we enhance these registry builders with these types of tools to speed up what they're doing?”
