From 3M Health Information Systems
Narrative text and NLP: Looking beyond your field of vision
Recently a colleague sent me a link to an article in the The New Yorker magazine titled “Tell Me Where It Hurts.” The article, written by Atul Guwande, surgeon and public health researcher, is about using incremental care in medicine. Instead of jumping to treat the symptoms, incremental care makes small changes in treatment or behavior, tracks the effect of the change, and then fine tunes as needed until the problem is lessened or gone.
The process reminds me of the continuous improvement cycle used in six sigma, SCRUM and other healthcare improvement programs, such as Deming cycle, Shewart cycle, PDSA (Plan, Do, Study, Act) or PDCA (Plan, DO, Check, Act or Plan Do Adjust, Act), the process of incremental care is used to make small changes and monitor the improvement. The cycle works like this: You “Plan” for the improvement by understanding the problem at hand and “Set” a goal for the improvement. Further, you decide to “Do” one small change and see if it makes a difference. You “Study” the impact of that change and then “Act” on the findings. You repeat the cycle over and over until you have met your goal. Sometimes the attempted change may not prove to be beneficial or it may demonstrate benefit, but now other areas require attention. This process is a great way to isolate a problem and make quick progress towards improvement.
In his New Yorker article, Dr. Guwande describes a patient who has suffered with terrible migraines for over 50 years. The patient explained to his doctor how migraines affected his life, with pain so bad that he would miss days of work, experience nausea and uncontrolled vomiting that often lead to dehydration, as well as visual disturbances even with his eyes closed, and an inability to sleep. The patient also explained the many things he had tried in order to lessen the frequency and magnitude of his migraines. Together, the patient and his doctor set reasonable goals and agreed upon journaling to record symptomatic observations moving forward. Additionally, they agreed and implemented one small change at a time and planned to meet every 2 months to review the impact (data) of these changes. Slowly, over the course of two years, incremental changes were implemented that greatly improved his quality of life. The key is that the physician and patient used data to achieve results. This data was captured in narrative text and reviewed for trends and patterns.
In hospitals and other healthcare settings, narrative text is a treasure trove of information that, once harvested, can provide clues to improve care and quality of life for all of us. The problem (until now) has been harvesting that data. In many cases, a manual process (requiring subject matter expertise) is necessary to review the documentation in order to find information and make inferences. However, manual review of documents is time and resource intensive. There is a significant opportunity for humans to interpret data rather than extract data.
Consider looking at the horizon with binoculars. The binoculars provide a targeted view when you know what you are looking for. However, you can only see as far as the magnifying power of the binoculars and you miss all the things outside of that field of vision. In health care, when doing a manual review of narrative text, you may only see what you are looking for and miss other information that may be related and significant. You need to be able to process a large amount of conceptual information in order to interpret the data for trends and patterns. When searching through text with Natural Language Processing (NLP), for example, there is an advantage over the human eye or binocular view, which can accelerate discovery and help gain insight into the complex inferences between clinical findings.
Take the concept of migraine. A search of clinical concept codes for migraine in SNOMED CT (Systematized Nomenclature of Medicine) returned over 120 codes. There are also related symptoms, medications, lab tests and even procedures that can be related to migraine. Using NLP to perform a concept search of text data, both structured and unstructured, combined with encoded data would give you a complete picture of migraine. The lessons learned from concept searches could reveal patterns that can influence treatment.
This is just the beginning of what merging clinical concept searching with NLP can do when combined with a knowledge model. Imagine conceptually identifying not just diseases but quality indicators, disease and device registries and social determinants of health for a given condition. You could have the clinical concept world at your fingertips.
Barbara Zellerino is a subject matter expert for Enhanced Natural Language Processing (eNLP) for 3M Health Information Systems.
What can you do with the healthcare data you can’t see? Find out.