Natural Language Processing, or NLP, is rapidly becoming a must-have capability for healthcare due to a myriad of factors, not least of which is the deluge of unstructured EMR data that must soon be accessible to patients per the Cures Act Final Rule.

By scanning a wide range of health care information and rapidly surfacing insights from unstructured text, NLP is a powerful tool to help payers and providers stay in compliance. But the true power of NLP goes far beyond compliance. We can also use it to augment risk stratification, better predict disease progression, identify gaps in care, and paint a fuller patient picture through social determinants of health – all pathways to better clinical care. This is the opportunity I’d like to focus on today.

Moving Beyond Electronic Health Records (EHRs)

Before the information revolution, it was impossible to understand the nuance required to tailor clinical care to the individual. Electronic health records, or EHRs, took us a big step forward, unlocking structured data that gave us insights into some – but not all – information needed to understand trends across patient populations. It is incredibly helpful to know, for example, diagnosis codes of specific populations at scale. But for a truly personalized, precise approach to clinical care, we need to know not just what is happening in our patient populations, but why.

NLP, the Foundation Technology for Deep Phenotyping

That’s why today, many providers are using NLP to perform deep phenotyping to better understand their patient populations and provide better care. The process involves gathering detailed information about disease manifestation, including granular information and clinical characteristics that are often only captured in written, free-text notes, and using it to better understand how a disease will progress in an individual, who is at higher risk, and which therapeutic approach has the best chance of success. The result is a bespoke, data-driven approach that’s optimized and tuned to account for each individual’s specific history and circumstances.

To effectively perform deep phenotyping in a practical and labor-efficient way, you need NLP. That’s because if we look at a typical EHR, only about 20 percent of the high-value information we need to accurately phenotype a patient is found in the structured format. The unstructured component of an EHR (including clinical letters, radiology reports, pathology reports and genetic test results) houses the deeper patient insights,  such as disease severity, treatment response, social determinants of health and more. By using NLP to unlock this rich context, providers can better understand their patients and connect the dots to forge better care pathways.

Understanding the Progression of Alzheimer’s Disease

Midwestern academic medical center (AMC) sought to examine common cognitive indicators of Alzheimer’s disease progression to understand what features are the drivers or predictors of more severe disease progression.

Through a partnership with Linguamatics, the AMC used a state-of-the-art NLP platform to build a set of pipelines that allow them to pre-process and clean EHR data from large cohorts, identify those phenotypic characteristics that are unique to Alzheimer’s disease, and build specific queries to extract discrete variables. By embedding the knowledge of what they need to extract from the EMR within the NLP (so called computational intelligence), the organization can include in their disease clustering models features from diverse and complex data such as neuroimaging studies and neurobehavioral tests. The NLP is able to extract these features with over 95% precision and recall, reliably adding features from free text to downstream predictive models.

These models are able to predict which patients will transition from mild to moderate disease, or from moderate to severe. By examining what features put patients into classes most at risk of transition between disease states, the AMC can intervene earlier and deliver better care.

Improving Diagnosis and Care for Patients with Aortic Stenosis

Kaiser Permanente sought out NLP for deep phenotyping with a different goal in mind. They wanted to improve diagnosis and care for one of the most common forms of valvular heart disease, aortic stenosis. Despite its prevalence, the optimal timing for follow-up for this condition is unclear and there is significant practice variation. Also, the natural history for aortic stenosis is outdated, relying on studies from 40 to 50 years ago. Conducting research for patients with this condition is difficult, and diagnosis codes are too broad to support the granular detail needed for better treatment decisions.

Kaiser Permanente leveraged Linguamatics NLP to extract key variables from its echocardiogram reports, constructing a complete picture of patients’ heart function and clinical details of their aortic valve disease. Once they created and revised NLP queries to achieve 95 percent positive and negative predictive values from the queries, they compared the accuracy of the NLP model to the standard codes-based approach to diagnosis. The Kaiser Permanente clinician-researchers found that within a large portion of patients who were identified as having aortic stenosis by the NLP algorithm, about one third would not be found by a codes-based approach, and about one-third of those with a nonspecific diagnosis code for aortic valve disease do not have aortic stenosis.

With their resulting database, which is the largest in the world for this condition, Kaiser Permanente can now examine the natural history of aortic stenosis and help update the outdated trajectory and rules currently used to risk stratify people. With more sophisticated risk pathways, they can now deliver more personalized and evidence-based care.

Reaping the Rewards of Precision Medicine

Whether in population health strategies, or personalized medicine programs – there is no doubt that the need for precision is absolute. To achieve this precision in a healthcare world of ever increasing volumes of complex unstructured data, NLP is essential in bringing this paradigm to life. We’ve long known that the information contained in unstructured content is valuable, but in the past, it simply wasn’t feasible to attain. Today’s technology puts rich insights at your fingertips and unlocks pathways that previously weren’t practical to pursue. From better clinical decision support to more sophisticated risk pathways and identification of gaps in care, deep phenotyping with NLP will ultimately help improve patient outcomes and operational efficiency while reducing cost. Don’t wait to embrace the full opportunities of NLP. Contact us today for a demo and take the first step toward optimized clinical care.