In a recent post on his Disease Management Care Blog, Jaan Sidorov, a physician with a keen eye for trends, speculated that natural language processing (NLP) might be used to pick up missing diagnoses from free text and perhaps even predict problems before physicians spot them. He cited a Mayo Clinic study that found that the use of an NLP program to scan free text in encounter records was nearly as accurate as lab tests in showing whether patients had the flu.
This is not a new idea. The University of Utah School of Medicine has been conducting studies of NLP for nearly a decade. But NLP is starting to become more capable, as shown by its growing use in computer-assisted coding. A VA study found that the use of NLP with free text identified post-operative complications more accurately than claims data did.
Meanwhile, Nuance Communications is working with the University of Pittsburgh Medical Center to apply NLP to electronic health records documentation, using IBM's Watson technology. The hope is that NLP will eventually be able to parse medical terms in free text to speed up data entry. Also, the researchers would like to be able to apply analytics to the data and generate "smart alerts" that can help providers improve patient care.
While the Mayo study focused on using NLP in biosurveillance to spot disease outbreaks early, Sidorov was intrigued by the possibility of employing such a system to detect diseases that physicians had not yet diagnosed. In addition, he noted, NLP-based analytics might be able to identify risk factors and "prospectively identify those persons at greatest risk for future complications, such as an avoidable hospitalization."
More immediately, he noted, NLP could be used to pick up diagnoses that have not been entered in problem lists--a continuing issue with EHRs. NLP is not the only way to do this; a recent study showed that a decision support tool could infer probable diagnoses from medication, lab and billing data. But only the NLP method can extract such data from free text.
Structured data in EHRs continues to be suboptimal for a number of reasons, including EHR design, lack of lab interfaces, and physician resistance to inputting data into point and click templates. Someday, when NLP can be reliably used to turn dictation into discrete data, physicians will be able to use EHRs more effectively, and it will be easier to measure their performance. Until then, however, the issues with EHR documentation are unlikely to disappear.
The idea of using NLP for predictive modeling and alerts, meanwhile, will continue to gain traction as researchers discover new ways to apply the growing power and speed of computers to medicine. Of course, computerized insights will never replace the intuition and knowledge of a skilled, experienced physician. But it would be nice if he or she had that extra edge. - Ken