Using Natural Language Processing to Crunch Your Clinical Data

For clinical applications, natural language processing can be used to search EMRs for information in both discrete fields and chart notes.

Natural language processing could be one piece of solving the EMR information overload puzzle.

"It's a necessary part of taking your EMR to the next level," says Walter Niemczura, director of application development at Drexel University's information technology department.

To mine clinical data, NLP scours electronic medical record systems for keywords and phrases related to patient care, generating information that can improve billing, efficiency, and population health initiatives.

"When we first started NLP and everything was on the research side, we talked about the time that could be saved doing research. Now, we're at the point where we can use NLP as part of an AI solution to improve care and improve the business," Niemczura says.

NLP can efficiently review physician notes from patient visits, he says.

"For example, a patient goes to our HIV clinic and the patient is billed for HIV, but the patient is not coded for hepatitis C even though they may have been prescribed hep C medications. Instead of manually reviewing every chart—no organization has that manpower—we can come up with a daily NLP task to see whether there was a missed billing opportunity," Niemczura says.

Automated chart review
 

Drexel clinical researchers used NLP to sift through 5,700 patient records for HIV and hepatitis comorbidity. "We reduced the number of candidates to 1,150," Niemczura says.

The project demonstrated the potential for efficiency gains from NLP, he says. "We were able to reduce chart review to 20% of the population. The original effort was five months, and we got it down to less than a week."

Using NLP to search chart notes was a key capability in the comorbidity effort, Niemczura says.

"If you were just looking at the codes in discrete fields, you came up with 677 patients, but there is more information in the notes. For patients who were HIV coded, the hep C patients could be absent in the codes but listed in notes. So, there were an additional 443 comorbid patients in the notes," he says.

NLP staffing
 

Niemczura says experienced IT staff should be involved in NLP projects.

"Unstructured data is not an academic query that you can learn in school—you learn from a lot of experience. So, the more experienced the user is, the quicker they can generate the end result."

IT specialists also can help integrate information from multiple computer systems and programs, Niemczura says.

"Having an informatics department involved is helpful because NLP technology is not just a standalone system. When there are billing opportunities or opportunities to improve patient care, researchers and clinicians don't have the ability to integrate the output from the NLP system into other systems such as the EMR," he says.

Lean approaches to NLP are feasible.

"You are better served with an informatics department, but it is certainly something that a physician could take the time to learn. Someone involved with research could learn it," Niemczura says.

EMR mining operation
 

The St. Louis-based Mercy health system's NLP initiative includes gauging the performance and outcomes of medical devices and searching patient records by disease symptoms.

"We were able to capture a lot of key cardiology from our notes. We put those results into a data platform that included all of the discrete data that we were able to get from our EMR and created a complete data set on a heart failure patient population," said Kerry Bommarito, PhD, manager, data science/performance analyst-enterprise at Mercy.

NLP allowed Mercy to collect a rich set of medical device information, she says. "We were able to show the life cycle of a heart failure patient to see risk factors for heart failure, medications, labs performed, date of heart device implantation, and outcomes such as ejection fraction results."

The project showed NLP has the potential to ease the EMR burden on physicians, Bommarito says.

"We could have asked physicians to document this information into discrete fields, but they already spend so much of their time documenting. Continuously asking them to enter more things into discrete fields takes away time they actually get to practice medicine and be with patients," she says.

Mercy has also used NLP to search for patient symptoms such as shortness of breath, dizziness, fatigue, and peripheral edema. From the onset, the project faced a linguistic obstacle, Bommarito says.

"There is no one routine way that a physician talks about a patient's symptom—everyone says the symptoms differently. Physicians will say, 'history of,' or 'patient does not have,' or 'patient denies,' or 'patient asked about.' There is a lot of work looking for patterns when finding symptoms like shortness of breath—physicians can abbreviate it to SOB or use synonyms like disnea," she says.

NLP eases the collection of EMR information, Bommarito says.

"There were things that we were looking for that we could not get out of discrete fields in our medical record system. We knew physicians were putting this information in procedural reports, discharge summaries, and patient progress notes. So, we are looking to use natural language processing to get information out of the notes and put it into a structured format."