NLP: An Evolving Dictation Opportunity
By Selena Chavis
For The Record
Vol. 29 No. 6 P. 16
Formative study findings suggest improved workflows for physicians.
The HIT movement has uncovered two truths about documentation best practices: structured documentation holds great promise for information sharing and analytics, and the patient narrative presented in the form of free text is a key component of optimal care delivery. While striking the right balance between these two modes of documentation is no easy feat, natural language processing (NLP)—the use of artificial intelligence to translate between computer and human languages—has emerged in recent years as a potential answer to this ongoing conundrum.
Further underscoring the NLP opportunity, a recent study published in the Journal of Medical Internet Research: Medical Informatics found that NLP-enabled EHR dictation has the potential to reduce the time required for documentation and improve usability. Considering a growing body of evidence that suggests EHRs have fallen short of delivering anticipated workflow enhancements, these findings could be good news for today's physicians who, by and large, are frustrated with EHR use.
"Documentation continues to consume a significant portion of physicians' workdays. If documentation becomes more efficient, physicians will be able to see more patients," says James Maisel, MD, one of the study's authors and chairman of ZyDoc. "From the hospital or practice's perspective, this increases revenues and profits. From a societal perspective, increased physician efficiency can slow the growth rate of the cost of health care."
Dave Kaufman, PhD, lead author on the study and an associate professor with Arizona State University's department of biomedical informatics, suggests that documentation has become an onerous process, one that physicians often view as an end unto itself as they labor to fulfill requirements and obligations associated with patient records. While EHRs are unquestionably promising, he points out that they are also immensely complex and suffer from poor usability associated with structured documentation. As a result, clinicians tend to prefer expressing themselves in narrative or free text.
"That [free text] is more difficult to analyze," Kaufman says, explaining that free text hinders EHRs from realizing their potential from an analytics and data sharing perspective because analysts have difficulty working with narrative, where important information can get buried. "You need a system that will parse and actually do NLP and recover the structure. EHRs can serve multiple purposes. They serve clinical purposes, but they could also be repurposed for quality assurance and research."
Acknowledging the promise of NLP and artificial intelligence for clinical workflows, Gilles-Andre Morin, chief technology officer at iMedX, expresses cautious optimism as he points out that "such technology is difficult to integrate in the physician's practice pattern without effecting significant changes to the physician's habits."
The standard input method of most EHRs requires point-and-click workflows, he notes. However, physicians hate to waste time, which is commonplace when using the EHR standard documentation method. While NLP may improve workflows, it typically requires voice input through front-end speech recognition, necessitating physicians become editors—a role they often deem more irritating than the time spent with a point-and-click EHR interface. "So adoption of front-end speech recognition and NLP in an acute care setting, where the cost of the documentation is not the responsibility of the physician, is likely to be met with resistance," Morin says.
The Study: A Deeper Look
"Natural Language Processing-Enabled and Conventional Data Capture Methods for Input to Electronic Health Records: A Comparative Usability Study" sought to evaluate the effectiveness of an NLP-enabled data capture method using dictation and data extraction from transcribed documents. Researchers considered documentation time, documentation quality, and usability of the method vs standard EHR keyboard-and-mouse data entry.
The formative study investigated four combinations of NLP and standard EHR data-entry methods, comparing a dictation-based protocol for structured data capture with a standard structured data protocol, as well as two hybrid protocols. Research encompassed 31 participants, including neurologists, cardiologists, and nephrologists, who generated four consultation or admission notes using four documentation protocols. In tandem, the researchers recorded the time on task, documentation quality (using the Physician Documentation Quality Instrument, PDQI-9), and usability of the documentation processes.
The study examined a method of EHR documentation by which the physician dictates, the dictation is transcribed or subjected to speech recognition, an NLP tool (MediSpaien) generates structured data from the transcription, and the structured data and text are inserted into the EHR.
Notably, the paper's findings suggest that EHR documentation methods using NLP-enabled dictation have potential for reducing documentation time and increasing usability while maintaining documentation quality relative to current mainstream EHR documentation practices completed via keyboard-and-mouse entry. "The NLP-based method was shown to require approximately 60% less time for documentation of a note than standard EHR entry using keyboard and mouse," Maisel says. "This difference can be explained by the faster speed of dictation relative to that of entering data using the keyboard and mouse, rather than by the involvement of NLP."
Maisel adds that the study revealed no significant difference between the quality of documentation generated using NLP-based methods and standard methods involving typing and using the mouse. "It is likely that by requiring that the author follow a more specific dictation template, or by improving the process by which the output of NLP is translated to an encounter note, one would be able to further improve the quality of documentation generated using NLP-based documentation methods," he says.
Finally, NLP-based dictation methods were shown to generate higher usability ratings than EHR documentation via typing and use of the mouse.
Kaufman emphasizes the importance of placing formative research in context. "Like in much of this kind of research, findings are always provisional," he says. "They usually are very much contextualized, but they feed ideas. And these findings suggest that this is a viable approach."
Morin agrees. "It is no surprise to me that new modalities involving the use of NLP and artificial intelligence technology have a positive impact on the effort associated with documenting clinical encounters," he says. "The devil is in the proverbial details. The big issue I have with this study is related to the fact that its results are the result of using the technology in a vacuum."
In tandem, Maisel points to a number of the study's limitations, including the following:
• The simulated, manual interface used in the study to move NLP-generated data into the patient's record is different from the automated interface that would be used in a real-world live setting.
• Physicians did not have an opportunity to review the documentation produced via the NLP-enabled dictation methods before they were finalized. In a real-world setting, they would.
• Physicians generated documentation for the study based on written test scripts about fictitious patient encounters rather than from evaluating a real patient by gathering information from various sources, including, for example, written records, conversations with colleagues and the patient, and imaging results.
• Three medical specialties were studied (cardiology, nephrology, and neurology). The sample sizes for cardiology and nephrology were small due to recruiting challenges, but all results were statistically valid.
• Because the study's methodology presented subjects with test scripts in a free-text format, it may have favored documentation methods requiring the physician to generate free text (ie, typing on the keyboard and dictation) and disfavored other documentation methods such as data entry via mouse.
Workflow Integration Opportunities and Challenges
Industry professionals agree that, although the concept of NLP-enabled dictation is worth further investigation, both opportunities and challenges exist to workflow integration.
Kaufman points to widespread HIT weariness that has resulted from a barrage of new ideas and workflow solutions coming on the scene in recent years. "Clinicians are continuously being introduced to this new feature or that new feature, or this new way of documentation that is supposed to improve their life and make work easier, and sometimes it fails to live up to that," he says. "They have become a little wary of the introduction of novel approaches."
Prior to tackling the full NLP payload interfacing challenge, Maisel suggests physicians need to buy in to the NLP-enabled dictation documentation method. In other words, implementation teams must convince them that the solution is faster—not only in theory but also in practice—and that it would not result in a deterioration of documentation quality. "This may be accomplished with an end-user study in one department moving the section level text and then expanded quickly," he says. "The major implementation barrier is the EHR company allowing this and extracting some fees."
Acknowledging that implementation challenges always accompany new workflows, Maisel also points to opportunities. He explains that physicians are eager to use dictation as a means for data capture, and they can delegate any remaining work on the document. This dictation model also provides physicians the convenience of documenting from anywhere at any time.
"The solution can be implemented with a smartphone app for dictation interfaced to the appointment schedule," Maisel says. "Telephone dictation and handheld recorders are also well accepted, so there is no significant training and quick adoption once implemented. After dictating, physicians have the option of delegating coding and document quality improvement to others before signing off on the document in the transcription system or after it is inserted into the EHR encounter."
To reduce the burden on physicians, Morin also envisions a workflow that enables use of front-end speech recognition and NLP to capture selective portions of the clinical encounter, but with the option to send the dictation over to transcription when self-edit becomes too cumbersome. "Such a workflow would have to be engineered for each type of EHR implementation and may not scale well from an NLP technology vendor perspective," he says. "But this would improve adoption of the new modality."
Impact on HIM
NLP-enabled dictation is good news for HIM, Kaufman says, noting that anything that helps recover structure from free text enhances the work of these professionals. "It's very difficult to do quality assurance or look at the efficacy of treatment costs when important information is buried in text," he says.
Morin agrees, noting that "The value from improved documentation has ramifications through other aspects of the revenue cycle management business, from improved collections and reduced denials to better operational metrics such as DFNB [discharged not final billed], quality, core measures, and others."
The large amounts of structured data that NLP-enabled documentation processes would generate may satisfy or help automate some work for HIM, Maisel says, pointing to federally mandated reporting, billing, population surveillance, alerts, and managerial reports. In addition, the generation of ICD-10 codes could provide significant labor savings for coding and earmark documentation that is inadequate for billing specificity.
What's the Cost?
Like any new initiative, the industry can expect an initial investment before realizing any return on NLP-enabled dictation. Kaufman points to software, training, and possible hardware costs. "When you implement anything new, there is always downtime. You have to allow for that," he says. "Staff has to be trained to a certain level of proficiency with the software. On the other hand, physicians do documentation all the time, so I don't think it would be too prohibitive."
In addition to the costs associated with bringing NLP to the market and implementation, Morin notes that the technology requires domain-specific training, meaning its usefulness in acute care settings is specialty specific. "There is a cost associated with training the engine, regardless of the underlying engine [rules-based vs machine learning], and this cost will likely have to be passed on to the clients," he says.
Maisel points to the initial and ongoing costs associated with the development and maintenance of an interface between the NLP and EHR systems. There are also the price tags of the NLP process itself and of converting dictation to text, whether by transcription or speech recognition. "It is about the cost of transcription plus a slight fee for the NLP processing and insertion setup and maintenance," he says. "The EHR may or may not charge for the interface. It is free with AthenaHealth and $10,000 for eClinicalWorks for a basic-level interface."
The bottom line is that this initial formative study offers enough positive conclusion for further research. Since the study, Maisel and his team are at different stages of implementing the workflow model in several hospitals and clinics. "We are working with several EHRs to enhance their ability to move more structured data into their systems," he says. "We continue to show the value of the NLP-generated data outside the EHR for both individuals and populations."
— Selena Chavis is a Florida-based freelance journalist whose writing appears regularly in various trade and consumer publications, covering everything from corporate and managerial topics to health care and travel.