Natural language processing, disease progression and population health

February 5th, 2018 / By Katie Christensen

Population health. The name pretty much says it all, right? Through the Triple Aim initiative, we are striving to improve the health of our overall population and make an impact on disease progression. This means fighting the ocean tide and preventing members from getting sicker, such as preventing the initial onset of diabetes or morbid obesity while working to stem progression and development of comorbid conditions (e.g. heart disease). The strategy is to find those patients who would benefit from earlier and appropriate interventions.

But do we know who these members are? Or will be?
Morbid obesity and diabetes are certainly oft-cited examples of conditions that are associated with follow-on complications, but clinical depression, which may go undocumented, also shows a correlation to disease progression.

This quantitative review suggests that depressive symptoms contribute a significant independent risk for the onset of coronary disease1

Suppose morbid obesity and clinical depression were clinically captured years earlier?
Both face different challenges in getting recorded in the medical record. BMI (needed to support morbid obesity diagnosis) should be readily available in the EMR in the facility setting, but is often not separately captured during an office visit. It is not uncommon for a professional claim to only contain the principal diagnosis.

Although there are guidelines for diagnosing depression, it is unfortunately not as straightforward as BMI since it relates to a person’s ability to actively function as part of the community. This is where natural language processing (NLP) may play a role:

Early work on diabetes outpatient records compared diagnosis by coding versus by NLP—NLP improved detection of depression diagnosis by almost a third…2

Why isn’t all of this technology in place today? What’s in the way?
Although NLP has been available, it has not historically been woven into the fabric of our coding dialogue. We are looking for the physician to document treatment supporting a condition and we seem to have challenges enough with physician notes in the EMR and coding queries. Although extensive effort is placed in the facility setting to ensure that all treated conditions are coded, the extent of HIM coding is not as pronounced in the professional setting with more limited office staff. Furthermore, NLP technology would need to be available. To exacerbate the issue, ICD-10 offers up to 70,000+ distinct codes that are used to capture the anatomic specificity of the disease. Given that diagnoses assignment is directly tied to payment, there is a reasonable concern related to upcoding activities. “Upcoding” refers to coding a condition on the claim despite questionable supporting documentation.

Shared savings and value-based programs
Thanks to shared savings and value-based programs, we have all the more incentive to fully capture patient risk and manage these members. With an industry move towards value-based payment, the reimbursement impact of a condition on a specific claim is offset by the overall performance on total cost of care over time. Additionally, shared savings arrangements generally have some risk adjustment methodology built in to measure actual spend against the disease burden of the member population. Reasonably enough, sicker members tend to utilize more resources and we expect them to be more costly. To this end, it is financially beneficial for all parties to fully capture the risk of their population, both in the short and long-term. In the short-term, we expect greater spend for sicker patients. Longer-term, if we capture, document and bill clinical depression or even lower back pain, perhaps we can prevent downstream clinical progression and even surgery, which would impact total cost of care (another part of the Triple Aim). Medical episodes offer an alternative insight into managing conditions once they are captured.

NLP can ease the current physician burden
Natural language processing offers an additional pathway to identifying the at-risk population. Physicians now spend “…33 percent of their net workday actually taking care of and interacting with patients. For every hour of direct patient care activity, two hours are spent on typing, data entry and paperwork.3
At the end of the day, it is in support of better health, lower cost and high quality that we fully capture the disease burden of our member population as early as possible. It will free physicians to manage these conditions, mitigating the need for surgery and/or risk of acute onset that would necessitate an ED visit or hospital admission.

Early identification. Active management. Patient-centric care.

Katie Christensen is a healthcare consulting manager within the Population and Payment Solutions group of 3M Health Information Systems.

Natural language processing meets healthcare

1. Wilson. L.R., Singal, B.M. “Do depressive symptoms increase the risk for the onset of coronary disease? A systematic quantitative review.” Psychosomatic Medicine, 2003 Mar-Apr; 65(2): 201-10.
2.  Geraci, J., et al. “Applying deep neural networks to unstructured text notes in electronic medical records for phenotyping youth depression.” Evidenced-Based Mental Health 2017 Aug; 20(3): 83-87
3. Benaroch, R. “How do doctors spend their days?” MedPage Today,