Home  |   Subscribe  |   Resources  |   Reprints  |   Writers' Guidelines

April 2014

NLP Shows Off Its Versatility
By Juliann Schaeffer
For The Record
Vol. 26 No. 4 P. 24

From CDI to research and patient engagement, natural language processing handles them all with aplomb.

Natural language processing (NLP), a technology that uses complex algorithms to find definition in free text, is proving useful to the health care industry as a whole, and it’s not just physicians and other medical professionals who can gain insight from the technology. New uses for NLP prove that potential benefits extend to medical researchers and even patients, and include helping boost clinical documentation improvement (CDI) programs, aiding chronic disease research efforts, and helping patients get more involved in their care.

A Boost to CDI Efforts
CDI isn’t new, but it’s new to technology, says Steve Bonney, executive vice president of business development and strategy for Records One, a circumstance his company and other vendors are aiming to change through platforms designed to help CDI specialists improve efficiency.

Without NLP, CDI specialists are used to manually checking records, with just basic search functionality at their disposal. “And it takes a lot of time and effort to do that,” says Juergen Fritsch, chief scientist and cofounder of M*Modal, adding that NLP technology can help automate the process.

When CDI specialists open a patient encounter, they’re likely looking at various electronic documents, such as the history and physical, consult notes, and an operative report. Without help from an NLP platform, specialists must manually search each document in every patient’s file to locate key terms. NLP can assume that task and complete it in much less time.

“One of the highest impact conditions that CDI workers are going after are heart failure cases, one of them being congestive heart failure [CHF],” Fritsch says. “Almost all CDI programs have their CDI specialist check for patients with CHF because you have to document a very specific aspect of CHF, whether it’s acute, chronic, or acute on chronic. If a document is determined by the NLP technology as not having the necessary specificity, it’s flagged for further review by a CDI specialist.”

A CDI specialist can open an encounter and review the NLP results to quickly determine whether any red flags are present. Is CHF mentioned and unspecified or is chronic kidney disease mentioned but not specified? What used to be a labor-intensive search now is completed electronically, allowing CDI specialists to concentrate on the more nuanced particulars of their job, such as determining whether the findings call for a physician query.

Bonney says the potential efficiency benefits to health care organizations are enormous. “Those improvements then lead directly to financial improvements,” he adds. “If you can review more cases, you should have better outcomes financially.”

NLP certainly isn’t error-proof, but Fritsch says even though the technology won’t find every minute detail, it’s still a tremendous tool “because it brings up more records more quickly and allows the CDI worker to find the information they’re looking for more quickly.”

“Anytime technology is involved, there’s an opportunity for error. I don’t think you can avoid that,” says Bonney, who believes the potential benefits greatly outweigh any downsides, particularly because CDI isn’t directly involved in patient care, where any risk is more paramount. “We’re not trying to replace people; we’re trying to give them tools to help them.”

Because there’s a shortage of CDI professionals, Bonney says health care organizations need the CDI department to be as efficient as possible. “The fewer people you have to do the job, the more important the technology becomes,” he says. “It’s a lot easier to train somebody in their job when you give them better tools.”

NLP also can aid ICD-10 preparations, Bonney says. “Having NLP should give you access to data that can help with physician readiness,” he says. “If you have NLP, you can generate reports that say which doctors aren’t adhering to the new documentation protocols as they’ve been trained, so they need to be trained again.”

“Knowing what the requirements are for ICD-10, NLP can tell you, for any particular disease that you’re documenting, whether you need to add more detail,” Fritsch says, noting that NLP also is an educational tool. “As an example, for fractures, physicians will need to document the laterality of the side on which it occurred. So if the NLP recognizes that a physician didn’t provide whether it was the left or the right knee or the left or right leg fracture, then it can flag that and bring it up as an educational remark.”

Indeed, NLP for CDI goes beyond simple documentation improvement on the back end. According to Fritsch, physicians can use the technology in real time to proactively fix documentation issues. “NLP in real time can really help close that gap so that physicians are documenting what they should be in the first place, which can also minimize the work on the back end that CDI specialists have to do,” he says.

“Don’t look at it from just the CDI specialist’s perspective or just the physician’s perspective, but see them as a team that needs to collaborate moving forward in order to get the best possible documentation,” Fritsch adds.

Chronic Disease Research
With its ability to mine the wealth of data contained in EHRs, NLP is finding a niche in medical research. While investigations relating to how NLP can help speed medical research still are in their infancy, its potential to reveal clues about various diseases is limitless, according to Jonathan L. Haines, PhD, a professor, the chair of epidemiology and biostatistics, and the director of the Institute of Computational Biology at Case Western Reserve University in Cleveland.

“EHR data has vast potential for research because of the depth and breadth of data collected,” he says. “One of the major hurdles is that the majority of these data is unstructured, mostly in free text. NLP is our way of generating structured data out of the unstructured data and thus making it useful for research.”

According to Haines, the longitudinal data in EHRs particularly is relevant for chronic disease. He, along with former graduate student Mary Davis, PhD, who now is an assistant professor at Brigham Young University, and other researchers at Vanderbilt University Medical Center’s Center for Human Genetics Research, recently conducted a study in which NLP in EHR systems was used to identify patients with multiple sclerosis (MS) and collect information on disease traits.

Although there already exists several fairly large datasets of MS cases that have been used for genetic studies, Haines says the detailed clinical data available on most of these samples are sparse, particularly for longitudinal data. “Thus we were motivated to mine the EHR to obtain these data,” he says. “The intent was to determine how much detailed data we could actually get and ultimately use in genetic studies. We were also motivated to determine the time and cost savings of using an existing EHR-based biobank to generate a dataset vs. taking the usual route of recruiting such participants de novo.”

Using NLP technology, researchers extracted eight clinical attributes from clinic notes, problem lists, and letters, including the following:

• clinical subtype;

• presence or absence of oligoclonal bands;

• year of diagnosis;

• Expanded Disability Status Scale and timed 25-foot walk;

• year and origin of first symptom; and

• medications.

How did NLP cull such information from unstructured data? “We did this by identifying key words and phrases, including misspellings, and excluding those in a questionable context—for example, with modifiers such as ‘possible’ or ‘questionable,’” Haines says. “By extracting up to 700 characters around the key words, we were able to obtain such context.”

From these extracted data, the researchers made several determinations, the first being that with the use of a fairly straightforward electronic algorithm, they could identify MS cases in an EHR with high specificity and sensitivity, along with well-matched controls.

“Second, [we determined] it is possible to obtain high-quality data using NLP methods, including such things as site of first symptom, progression of the symptoms and disease, and even quantitative data, such as the timed 25-foot walk,” Haines says. “Third, although not surprising, it is clear that there is substantial heterogeneity in what data you can recover from any individual record. So although we ended up trying to collate seven or eight different variables, very few records had all this information available.”

While Haines and the other researchers largely knew the information and data they expected to find, NLP did turn up one particularly surprising finding. “In the case of the timed 25-foot walk, it was often recorded both in the free text and in a structured field,” Haines says. “We were somewhat surprised, but very pleased, to see the very high correlation between the two.

“Personally, I was a bit skeptical about how much of this detailed clinical information we could extract from the EHR, but it worked quite well,” he continues, noting his hope that improved algorithms will be able to catch even more details of disease clinical expression and progression.

For patients already diagnosed with MS, as well as those who have yet to be diagnosed, Haines says NLP holds great promise. “This should speed research into MS, as we should now be able to access a very large amount of existing MS data sitting in EHRs in many different institutions,” he says. “This is more of a proof of concept that such data can be captured, but with these kinds of data, we may be able to better focus our research on different subtypes of MS, and on things such as progression.”

And this isn’t a lone case—or disease—that can be aided by NLP. According to Haines, similar approaches can (and are) being applied to many other common chronic illnesses, including diabetes, cardiovascular disease, and even adverse drug responses.

He notes that the options in this regard (and the potential for patients who stand to benefit from improved diagnosis and treatment protocols) are seemingly endless. “Finding ways to open up the EHR will speed research for virtually anything that’s recorded,” he says.

Engaging Patients
Helping health care professionals, even researchers, do their job more efficiently is one thing, but could NLP technology help patients, too? Next IT believes it can and has developed a virtual assistant platform that uses NLP technology in an effort to engage patients and get them more involved in their own care.

According to Victor Morrison, Next IT’s senior vice president of health markets, the impetus for this tool, called Alme for Healthcare, stemmed from the challenges currently facing the health care system—as well as what could be more trying times ahead.

“It’s no secret that our health care system is facing some steep challenges right now,” he says. “Due to an aging population as well as an increase in overall coverage caused by the Affordable Care Act, our health professionals are being stretched to their limits. The Association of American Medical Colleges has warned that the increase in demand for health care will lead to a doctor shortage of more than 90,000 by 2020. There’s no way we can combat this shortage by sticking to the status quo.”

Seeking solutions that provide proactive, preventive treatment rather than continuing to lean on costly emergency department care, Next IT views virtual assistants as one solution to help patients themselves alleviate some of that pressure while improving care.

Next IT is well versed in the virtual assistant realm, having built platforms utilized by United Airlines, Aetna, and the US Army. But with Alme for Healthcare, the company has designed a platform specifically for health care systems, and it’s hoping to open the door to improved patient-clinician interaction, leading to improved outcomes, greater patient satisfaction, and lower costs.

“Patient adherence, or the lack thereof, is one of the biggest sources of waste in our health care system, costing up to $290 billion each year,” Morrison says. “By providing an interface patients can interact with daily on their smartphones, tablets, or computers, Alme for Healthcare’s technology helps patients stick to the treatment plan they and their doctor have laid out. Meanwhile, doctors can be updated with patient data in real time, so they can spot problems long before the next scheduled appointment.”

The technology is built on three core pieces: a comprehensive patient ontology, goal-based tracking and conversational engagement to help patients maintain health plans, and interactive illustrations to help patients manage at-home tasks. “These parts all come together in a conversational interface that can be used in a wide range of health-related applications,” Morrison says.

Using NLP algorithms developed specifically for health care–related environments, the virtual assistant can recognize a wide range of medical concepts through conversational language. “It can help record daily treatment progress, missed injections, unexpected symptoms, [and] mood, and even call 911 if necessary,” he says. “This information can be used by the patient and their physician to track progress and overall success for their treatment plan.

According to Morrison, patient information is protected because the Alme for Healthcare software lives on each customer’s own network. “They have total control over the security and safety of their patients’ information, just as they would with any other part of their network,” he says.

One of the technology’s most interesting applications involves chronic disease management. “Take type 2 diabetes, for example,” Morrison says. “Managing diabetes can be a very exhausting experience at times, both physically and mentally. Alme for Healthcare–powered virtual assistants can make this easier by helping patients track injections, find optimal injections sites, and encourage them with positive interactions.”

Morrison says NLP-driven tools have great potential to improve both patient outcomes and quality of care because they use the as-yet-untapped power of the patients themselves. “With doctors and other caregivers already stretched as it is, bringing patients in as equal partners in the treatment process is one of the most critical goals for health care providers right now,” he explains. “Doctors, nurses, and clinicians can’t watch over patients day and night, but by empowering patients to make informed decisions, virtual assistants can help fill the void in between scheduled visits. The impact this can have on patient outcomes is tremendous.”

Looking long-term, Morrison sees many more ways that NLP can help connect patients. “The potential for NLP and virtual assistants is truly fascinating,” he says. “As Alme for Healthcare and other NLP-powered assistants become increasingly advanced, patients will take a more central role in their health than ever before. By linking up virtual assistants to wearable sensors, genetic data, health history, and more, patients will gain a complete, real-time overview of their health.
“Someday soon, it will be the patient who notifies the doctor that something is wrong.”

In Summary
CDI, research, and patient engagement are just three ways that NLP is assimilating into the health care arena. While the technology still is new to patient care as a whole, and its use is nowhere near mainstream just yet, Bonney urges health care professionals and vendors to strive toward using NLP in even more ways for an even greater impact. “NLP at its core is creating structured data,” he says. “Now that you have structured data, what are you going to do with it?”

The options, particularly as they relate to health care, still are unfolding.

— Juliann Schaeffer is a freelance writer and editor based in Alburtis, Pennsylvania.