One organization’s use of natural language processing (NLP) has increased efficiencies, including more accurate data management, leading to the delivery of better care quality and improved billing practices.

NLP is a subfield of linguistics that combines artificial intelligence with computer science and information engineering to optimize the interactions between computers and human language.

In particular, it is used to help program computers in a way that will enable them to more effectively manage large amounts of data in the form of natural language rather than computer coding.

This approach has significantly improved the value of the electronic health record (EHR) at Drexel University College of Medicine in Philadelphia, says Walter Niemczura, director of application development. Drexel originally began using NLP to review unstructured notes within the EHR to identify patients with certain characteristics to improve healthcare research and patient care.

The new approach was needed, Niemczura says, because prior to implementing the NLP technology, researchers had to review 5,700 charts manually to find needed information.

With NLP, they reduced manual chart review to only 1,150, just one in six charts. That is a reduction of more than 80%.

Further refinement of the process is nearly eliminating the need for manual reviews. That reduction saves Drexel researchers and staff significant time and related costs, Niemczura says.

The NLP also yields a more complete analysis of the information within the EHR, which in turn produces more accurate analysis and decisions, he says.

Drexel has leveraged this accuracy for various use cases, such as patient trials on the research side, enhanced quality of care initiatives, and improved billing due to better data management, Niemczura reports.

“The original intent was to use it preparatory to research, searching provider notes for different research opportunities like comorbidities. It’s been very successful at that and a significant time saver,” he says. “Our first project was one that was done manually, with residents spending three or four months going through more than 5,000 patient records looking for patients with HIV and hepatitis C. We did it again using NLP. We discovered 1,100 patients that met required chart review for those criteria almost immediately.”

Searching Notes for Information

The difference from standard computer searches is that the NLP can search clinician notes for mentions of certain words or phrases that suggest the document is relevant, whereas other searches may be limited to specific ICD-9 codes or similar data points, Niemczura explains.

“We can say, ‘Here are the people who have ICD-9 codes for HIV and hepatitis C, but also here are the patients who have an ICD-9 code for HIV, but we found mention of hepatitis C in the notes field, and vice versa,’” he says. “We also can find those in which neither of the ICD-9 codes are assigned but one or both of them are mentioned in the notes.”

The NLP technology also highlights the occurrence of those keywords in the notes so that researchers do not have to scan the entire record to determine the relevance of the words to the project, Niemczura explains.

“We’ve been exploring using it on the operations side for opportunities to address missed billing and to update the medical record. With the HIV and hepatitis C comorbidities, we’re looking for patients in the record that show we billed for HIV but discussed hepatitis C during the visit,” Niemczura says.

Those records are tasked to the billing office to review the item for a possible missed opportunity, but that also is an important opportunity to update the medical record, he says.

“How many providers are reading through all the previous notes? If it didn’t make it into the problem list, it’s not really at the forefront for them at the next patient visit,” Niemczura notes. “That has a big cascading effect. Having the problem properly documented helps avoid lawsuits in which they claim you shouldn’t have missed this in the notes, and it just improves the care you’re able to provide further down the line.” That improved documentation also can improve an organization’s data warehousing efforts, Niemczura suggests. He compares the application of NLP to patient records to the editorial process found in published medical journals. Normally, a patient record is just a one-pass document in which whatever the clinician creates is accepted and filed as the permanent record, he explains.

“With the addition of NLP, it kind of forces an editorial process in which the artificial intelligence reviews the record and looks for opportunities to improve it, to make it more accurate,” Niemczura says. “The information is there already, so it’s not like you’re adding anything or altering the factual elements of the record. This process can help make that patient record a more meaningful depiction of what the physician recorded from that encounter.” The use of NLP also can identify frequent insufficiencies in patient documentation and provide an opportunity for education of clinicians, Niemczura adds.

Consider Technology Demands

The implementation of the NLP technology required a new server. Drexel’s EHR server is located in Milwaukee, so the school considered locating the NLP server there, too, or locating the NLP server in Philadelphia and transmitting the data between the two cities.

That second option would have required acquiring additional bandwidth for the data transmission if the records were updated in real time.

Leaders decided on placing the server at Drexel but relying on only a monthly backup for the data. That saved on the bandwidth expense because the first use of NLP was for research, and the researchers did not care if the data were only one month out of date.

“However, when you’re looking at billing opportunities, you can’t be a month behind on data. We started to leverage the application interface of our EHR so that we could have daily builds of files,” Niemczura says.

“That allows us to have everything here for the current day’s notes,” he continues. “At night, we can build and index, review it, and process it. For the infrastructure and getting the data, you have to see what your needs are to address that.”

Query Formulation Key

Developing the proper query for NLP is crucial to making the results useful, Niemczura stresses. Drexel researchers initially found that the best approach was to look at a sample of 500 records to identify query terms that would apply to their research needs, then determine how many of those records would have been flagged under those terms.

“That validates their process to show that for every 500 records, we expect 20 patients to be found. If you run the NLP with that query and get about that rate of records found, you know your query is on target,” Niemczura explains.

“If you’re not getting that return from the query, then you need to refine the query more,” he continues. “If you’re getting too many false-positives, you need to tighten the query to avoid that.”

Niemczura thought that could be improved, so he sought a way to set up the NLP technology to perform the first pass without the manual review.

In every case, the NLP outperformed the researchers and found more relevant records. The computer was more expansive in its search and more diligent. “The technology is finding terms, sometimes archaic terms, that the researcher was not looking for but which were in fact relevant to the search. The program is also wonderfully repetitive,” Niemczura says. “It has no fatigue. It has no distractions. It can do a better job.”

Repeated cycles of the NLP queries yield a nearly 100% level of confidence in the results, Niemczura says.

Nevertheless, it is important to review any instances of false-positives or overlooked records to ascertain why the NLP did not assess them properly, he adds.

Staff, Disk Space Concerns

Niemczura also emphasizes the need to check with end users to assess their opinion of the data and its applicability. It is important to avoid alert fatigue, he says.

Drexel absorbed the NLP technology without hiring additional staff, but Niemczura says some large-scale healthcare organizations might need to dedicate a new employee to managing such a system. The return on investment probably would justify the additional staff person, he says.

Disk space also is a concern, because the search index is a copy of the organization’s data. “This is an enterprise application that you have to devote resources to. It’s not just an app that you can pop onto a researcher’s PC,” Niemczura says.

“It’s on the scale of bringing in a new EHR, but it does require some investment,” he continues. “I think most organizations will be happy with the result.”


  • Walter Niemczura, Director of Application Development, Drexel University College of Medicine, Philadelphia. Phone: (215) 895-2000.