How NLP can help with at-risk patients, SDOH and pop health

Photo: Linguamatics

The new Centers for Medicare & Medicaid Services Interoperability and Patient Access final rule requires the interoperability of full-text medical records.

“The Interoperability and Patient Access final rule put patients first by giving them access to their health information when they need it most, and in a way they can best use it,” according to CMS. 

“This final rule focused on driving interoperability and patient access to health information by liberating patient data using CMS authority to regulate Medicare Advantage, Medicaid, Children’s Health Insurance Program and Qualified Health Plan issuers on the Federally-facilitated Exchanges.”

There is, of course, a massive amount of unstructured text in electronic health records. As a result, healthcare provider organizations are in need of a way to plow through all that text to comply with the CMS rule.

One technology more providers are turning to is natural language processing, a form of artificial intelligence. NLP can do that by plowing and turning free text into structured text, making it much more accessible and usable for end users – and patients.

Dr. Calum Yacoubian is associate director for healthcare strategy at Linguamatics, which develops a natural language processing-based artificial intelligence platform. 

He offered readers a brief primer on NLP, and explained how the technology can help clinicians identify at-risk patients, how it can help clinicians and health IT leaders uncover insights into social determinants of health and advance population health efforts.

Q. Please explain, in layman’s terms, how NLP works with unstructured text.

A. Think of NLP as the ability of computers or machines to understand language as it is used and meant by humans. It is not just searching for keywords, but identifying those words and understanding their meaning and the context they are written in.

Over the past decade or so, NLP has become something that is much more mainstream than many of us may realize. Take products like Siri and Alexa – these are examples of “machines” using NLP to understand the language of humans.

In healthcare, clinical NLP adds the nuance and peculiarities of clinical language into the mix – giving the machine the ability to decipher the unique ways clinicians document clinical histories of their patients. A simple example is the ability of NLP to understand negation.

Take smoking history for example: The statements “she is an ex-smoker,” “she denies smoking” and “she smokes 10 cigarettes per day” all contain very similar words but have three distinct meanings. NLP picks up on these differences and categorizes each statement correctly.

The big benefit of NLP is that it can do this “reading” of the text at great speed and scale. Modern NLP systems are able to analyze millions of documents per hour without fatigue.

Given that no clinician could do this, NLP can be a great companion for medical teams looking to synthesize insights from large volumes of complex medical records in an effort to better serve their patients and improve health outcomes.

Q. How can NLP help clinicians identify at-risk patients?

A. The first thing to be aware of here is that a great many risk factors for disease, disease severity and disease progression are features that either cannot be coded or are rarely coded because they do not impact medical insurance reimbursement through Medicare, Medicaid or private healthcare coverage.

These signs, symptoms and social determinants of health are captured in medical records by clinicians, but rarely translate to what exists in the structured data. Therefore, we know there is information in the free text that is useful in identifying risk. Knowing this – and with an understanding of what NLP is – we can begin to see the many ways that NLP can help clinicians identify at-risk patients.

First, let’s consider a case where risk factors are known for a disease. NLP can be configured to search the medical records for those known factors and determine which patients have these features present.

The more abstract but exciting area where we want to use NLP is to help determine patients at risk because of unknown risk factors. In this instance, we can use different techniques in NLP to tag the records with different features using different parts of the NLP process.

For example, NLP can determine which words are diseases, procedures, signs or symptoms – known as “entity recognition” – and then the technology can link those entities to known ontologies or terminologies to build a rich feature set of potentially predictive features.

These features then can be fed into other clustering or machine learning pipelines alongside standard structured data to determine potentially at-risk patients. This use of NLP to create features for machine learning is a powerful combination.

Q. How can clinicians and health IT leaders glean insights from free text data to uncover SDOH?

A. As mentioned, information on patients’ social determinants of health remains poorly coded in most electronic health records. Despite this, medical professionals are trained to capture a complete history from their patients, which includes information on that patient’s social history.

Through either dictation or direct entry into the EHR, this information exists in abundance within the free text of patients’ medical records. Knowing the information is there means we can use the exact same techniques to sort and organize the records for SDOH as we do for signs and symptoms.

Let’s take a very straightforward example. A clinician documents the following for an elderly patient at admission: “This gentleman was brought in by ambulance having been found lying outside his home. He lives alone, his wife having passed away three years ago. He normally uses a Z frame and cannot drive.”

This kind of documentation is standard in healthcare, and nine times out of 10 it remains hidden in the notes section of the EHR.

With NLP, this statement can be transformed into structured data that reads: “Social Isolation; Widowed; Walks with assistive Device; No transportation.” Having this information in a structured format against not just one patient but an entire population means providers can better identify those patients at higher need and higher risk.

Q. How can NLP help advance population health efforts?

A. While the last example on SDOH is a great illustration of how NLP can impact population health, we don’t have to focus on SDOH to see how NLP can help with population health initiatives.

Our understanding that the sickest in a population account for the majority of healthcare spending is not new. This is why the challenge has always been to identify and manage patients earlier and when their respective disease is less severe, so they do not progress to more resource-intensive needs.

As mentioned earlier, the parts of a patient’s clinical history that can inform this are rarely captured in structured data, so stratifying the population appropriately has posed a great challenge to healthcare organizations. NLP can provide a reliable, configurable and transparent way to enable more accurate diagnosis capture, and therefore better understanding of the true burden of disease.

Take a hypothetical example of congestive heart failure. By looking at only the structured data, we can gather some information on the population impact of this disease – but not at a level granular enough to impact population health efforts.

Using reimbursement billing codes, we can see there is a population of 1,000 patients with CHF, but we cannot discern their symptoms, their New York Heart Association classification or their severity. If they have been hospitalized, it is very difficult to tell if worsening heart failure has been a leading cause of that admission.

If NLP is applied, we not only potentially can identify a larger cohort of heart failure patients, but also capture information on their disease severity (for example, complaints of dyspnoea, leg swelling and difficulty sleeping) as well as hospitalizations due to worsening heart failure.

This can lead to a better understanding of which patients are at risk and where care gaps need to be plugged. It is one way in which Kaiser Permanente is using NLP – as published in JAMA last year.

Twitter: @SiwickiHealthIT
Email the writer: [email protected]
Healthcare IT News is a HIMSS Media publication.

Source: Read Full Article