Menu
Close
Home   >   Blog   >   Natural Language Processing in Healthcare: Turning Promise into Change

Natural Language Processing in Healthcare: Turning Promise into Change

by William Kinsman, Senior Manager of Product Innovation, on November 25, 2019
William Kinsman, Senior Manager of Product Innovation

If you’re still waiting for natural language processing to make an impact in the healthcare industry, a quick look around should tell you that artificial intelligence in healthcare — inclusive of natural language processing — is doing more than just promising change – it’s driving change. More than 100 million clinical records are reviewed each year in the United States. Those records contain data – lots of it. Encounter data, pharmacy data, patient histories, social determinants of health data – all telling the story of a patient. Some of it is in electronic form; some of it is free text. All of it is vital to understanding the patient’s journey. And natural language processing is perfectly aligned to quickly and efficiently deliver the insights healthcare organizations need to provide the best quality of care to patients, while also reducing total cost of care.

Today’s natural language processing technologies alleviate much of the burden of manually recording coded conditions within electronic health records for providers, improving medical record review accuracy, and streamlining medical record review for payers and health systems thus reducing the amount of time and resources spent. For example, computerized clinical decision support systems (CDS) — dynamic tools that leverage machine learning — help physicians provide better informed care; prompt providers during various points of care to order medication, schedule healthcare screenings, vaccinations, or other important health-related needs; and help provide easily accessible health-related information at the point and time that it is needed.

Without natural language processing, CDS as a tool would be far less useful. In a report in the Journal of Biomedical Informatics about Biomedical Natural Language Processing, researchers Dina Fushman, Wendy Chapman, and Clement McDonald note that “much of the data that supports CDS is textual and therefore cannot be leveraged by a CDS system without natural language processing.”

Putting this theory to test, Dr. Dominik Aronsky and his colleagues studied the usefulness of natural language processing for a CDS system that identified community-acquired pneumonia in emergency room department patients; the results showed that the performance was significantly better with the natural language processing output – once again proving the value add of the software. In the August 2019 issue of the npj Digital Medicine journal, Doctors Trishan Panch, Heather Mattie, and Leo Celi discussed the value add of AI in healthcare, stating:

“The rapid explosion in AI has introduced the possibility of using aggregated healthcare data to produce powerful models that can automate diagnosis and also enable an increasingly precision approach to medicine by tailoring treatments and targeting resources with maximum effectiveness in a timely and dynamic manner.”

Today, natural language processing is also able to assist payers in improving coding guidelines for their associated providers, often reducing overall abrasion between the two and improving risk score accuracy and quality reporting. While this is a huge benefit to payers, there are still instances when payers have difficulty integrating the tool with their current system processes, resulting in unstructured medical notes and narrative text, ultimately causing a delay in the medical record review (MRR) process. Appropriately leveraging natural language processing technology in the healthcare industry remains complex, and the right technology can be tricky for a variety of reasons.

7 Ways to Evaluate Natural Language Processing Technology

Language is one of the most dynamic and difficult challenges facing the use of machine learning in healthcare, as how people speak tends to differ geographically, in different disciplines, and even over time. This is especially true within the healthcare sector, as new drugs are created and our understanding of the human condition evolves.

However, there are multiple ways of evaluating the performance of the use and value of natural language processing technology in healthcare:

  1. Seamless integration – Natural language processing algorithms are already complex in nature, so having software capable of seamlessly integrating with current tools helps to augment the MRR process, saving both time and resources.
  2. New features – Does the software permit you to do things that you simply could not do without it? Natural language processing is unique in that it can read infinitely faster than a single human can, querying for data that would otherwise need to be manually identified. Sophisticated models have probabilistic output, where the computers indicate their confidence that they have found a target.
  3. Qualitatively – Does the software perform well when given to a team? Does it make sense and interact with their workflow? We don’t expect these tools to completely replace human interaction in the short run, but if they make tasks easier or produce higher returns for the same task without them, it is worth the investment.
  4. Quantitatively – Data scientists use numerous tools to monitor and understand the performance of any algorithm. For machine learning or deep learning in healthcare, one of the more popular tools is “precision and recall,” which roughly describes the percentage of time the model is right, as well as the percentage of time the model overlooks notable events it should have caught. A data scientist at your company can help break these down.
  5. Transparency – Though difficult with deep learning, being able to know why a computer is making a specified decision for someone is important, although vendors are less likely to share this information. Confidence that they at least have the ability to do so with examples, however, can give your team greater confidence the vendor is technically sound.
  6. Sustainability – Natural language processing models in today’s world are almost always “batch” trained. This means they stay static unless they are retrained, as opposed to “online” training where models continually learn. Batch training in healthcare decisions is a significantly safer choice because online models can sometimes learn from incorrect data and potentially degrade in performance over time if not monitored in real time by an expert. Knowing that a company has a plan in place to retrain, update or monitor their model is the best possible hedge for the future.
  7. User Interface – It is important to have natural language processing technology that allows users to manage all their medical record reviews in one place and provides manager-level oversight to assign and review medical records as necessary. According to Carol Friedman, a Columbia University scientist and biomedical informatician, natural language processing in healthcare and biomedicine can be leveraged in a multitude of ways:

“For quality and administrative purposes, NLP can signal potential errors, conflicting information, or missing documentation in the chart. For public health administrators, EHR patient information can be monitored for syndromic surveillance through the analysis of ambulatory notes or chief complaints in the emergency room.”

As natural language processing continues to evolve, so will its impact on the healthcare industry. But those organizations embracing the technology and choosing to evolve with it — rather than wait for it to realize its potential — will be the organizations that truly differentiate themselves from their peers.

Subscribe to Inovalon's Insights Blog to receive an email alert when new blog posts are published.
© 2019 Inovalon. All rights reserved.