Since 2002, researchers have attempted to assess the impact of social factors on health outcomes. Recent studies estimate that traditional medical care accounts for only 10-20% of an individual’s health, while the remaining 80-90% is due to SDoH.(1)
Regardless of where you sit on the political spectrum, I think we can all agree on some fundamental facts like the rising costs of healthcare and cost of living, economic disparities, mental health concerns, food insecurity, and shortage of housing, just to name a few. These concerns all fall into a category called social determinants of health (SDoH). Given the power this category has on healthcare costs and outcomes, everyone in healthcare is trying to figure out how to address these "social determinants."
A Closer Look
The goal is to gather the right data, extract the important information, and determine how and whom to help in a way that can facilitate long lasting, positive outcomes. At Health Language, we are sifting through the noise to extract valuable information that can be leveraged to support analytics and care programs. Currently we are focused on tackling healthcare’s noisiest data, the clinical notes.
There is a fantastic amount of information in clinical notes, and from the intake notes to the discharge summary we have found these notes to be rife with social determinants of health information. While this information can be used in a variety of ways, we are interested in learning whether the data can help us predict readmission risk.
While our experiment was initially driven by curiosity and innovation, our results were quite promising. Using our advanced Clinical Natural Language Processing (cNLP) solution, we unpacked a series of clinical notes, explicitly searching for 12 SDoH categories (2). By tuning the solution with our proprietary cNLP Lexicon, which includes content libraries, synonyms, misspellings, and provider and customer-friendly terms, we were able to look at things like employment, behaviors, income and social status, and physical environment.
Once insights were identified, our cNLP solution was able to normalize the results by codifying the extracted information to appropriate industry standards such as ICD-10 and SNOMED CT. From there, we grouped the outcomes and scored the results to help us assess the risk, and then further refined our results by analyzing the individual’s subsequent ER visits.
The outcome was surprising and informative. Here we saw the power of cNLP when applied to a real-world problem and significant healthcare challenge. At this point we can’t help but think of all the ways this data could be used to improve the healthcare delivery model and ultimately improve patient care.
Could this insight help the ER staff set up educational processes as part of intake? Could this data help determine how to follow up with the patient? Could the data be used to coordinate telehealth services? Could it tell us who might need economic assistance to pay for medications to encourage adherence? There are numerous scenarios where this data could be used to help reduce the risk of readmission within 30 days of discharge.
While our experiment proved to be very successful, we want to do more, learn more, and ask more questions:
- How does this help providers? How does this help payers?
- Could we help with Acute Myocardial Infarction (AMI)? Or maybe target Chronic Obstructive Pulmonary Disease (COPD)? What about Heart Failure (HF) or Pneumonia (2)?
- Who do we partner with to test this further?
- How does this affect what others are doing already?
- How can we augment those efforts with our capabilities?
- How do we integrate more SDOH?
These are just a few of the questions we are eagerly and enthusiastically pursuing, and actively in the process of answering.
Check back for more updates and news on our progress. Feel free to comment or reach out directly if you’re interested in participating in this exploration!