April 2017

Is Speech Recognition Viable in the ED?
By Julie Knudson
For The Record
Vol. 29 No. 4 P. 24

A small study explores the incidence of errors in this harried environment.

There's a lot going on in emergency departments (EDs), most of it centered on saving lives. Documentation may rank pretty low on the staff's list of priorities, but that doesn't diminish its significance. To make note-taking more convenient, some hospitals have turned to speech recognition for help. Is it effective?

Scott Weiner, MD, MPH, attending physician and associate director at Brigham and Women's Hospital Department of Emergency Medicine in Boston, describes the adoption of speech recognition software in the ED as a game changer. "We love it. I can't imagine going on shift without using it," he says. But when errors began popping up, Weiner says the team wanted to dig deeper into the causes. "Occasionally there were errors that were ambiguous, and we didn't understand the meaning from the context," he says.

In 2016, the University of Colorado and Brigham and Women's Hospital published "Incidence of Speech Recognition Errors in the Emergency Department" to explore more about the mistakes that had entered the EHR system through the ED. How many errors had actually been introduced? Were they trivial, or did they have the potential to impact the quality of care or create medical liability?

Weiner and his team intended to find out the real-world incidence and impact. "We did the study because we wanted to quantify our observations," he explains. A small sampling of 100 notes, each taken from the first half of 2012, was selected and reviewed for the pilot study.

Weiner says the researchers anticipated finding some errors but were surprised to find that 71% of the notes contained at least one error. "The majority you could understand the meaning from the context, so there were no adverse clinical implications," he says. However, some of the errors were of a more serious nature. In all, 15% of errors were ranked by the team as critical, meaning they could have affected providers' decisions about patient care.

Digging Into the Study's Findings
As hospitals seek more cost- and time-efficient ways to meet documentation requirements, the use of speech recognition in EDs is increasing. To have the technology perform at its peak, there are several steps organizations can take to avoid errors such as those found in the 2016 study.

The fact the study included only dictations by physicians whose native language is English must be taken into account, says Sheila Guston, CHDS, AHDI-F, president of the Association for Healthcare Documentation Integrity (AHDI) and supervisor of HIM document integrity at Spectrum Health in Grand Rapids, Michigan. "In a more random sampling of providers with a blend of dialects available, they may have seen an even higher error rate," she notes.

Problems are amplified when ED physicians don't have time to read through and refine their speech recognition results, Guston says. "Without proper editing of results, the system will continue to misrecognize what they're saying," she explains while noting that only one facility and one speech recognition software provider (Nuance) were part of the study. A sampling that encompassed a broader provider base and additional software platforms may have yielded different outcomes, Guston notes.

Hannah Zook, RHIA, director of health information services at Terra Nova Transcription, wonders how the mistakes deemed critical affected the organization. "Was patient care compromised?" she asks. "Was billing? Was there something incorrectly documented that could have caused abstracting errors?" From patient safety to reduced reimbursement, errors that persist within the EHR can have wide-ranging impact. "I'm interested to know what other items could have trickled down into the revenue cycle," Zook says.

How the speech recognition software was employed also interests Zook. "Did they sit at a station with the EHR or did they dictate directly into the system right there with the patient?" she asks. Background noise and other common distractions in the ED environment may explain some errors, but if entries were made elsewhere then other factors might have influenced the study's results, she says.

Tim Ruff, vice president of solutions management at M*Modal, says the study is part of an important line of research that may help clinicians, coders, and others in the hospital environment understand potential quality issues in clinical documentation. However, he notes the data used in Weiner's study were collected in 2012. "Speech recognition technology has evolved considerably in the last four years with next-generation speech engines in the market," he says. Accuracy and overall performance have improved significantly in the intervening years, as has the ability of speech recognition systems to continuously learn and refine their analysis functionality, he says.

While the potential for the errors found in the study to cause actual harm was of interest to several experts, Regina Kraus, MSA, RHIT, CPHQ, an AHIMA-approved ICD-10-CM trainer, of the Kaplan University School of Health Sciences, is curious to know more about how that trend appeared to researchers. "Were they just lucky?" she asks about the lack of harmful consequences. "Or was that the norm?" The study's sample size also piqued Kraus's interest. "It was just 100 notes, which I thought was a little small," she says, noting that the number of visits at a large academic hospital may lend itself to a review of a greater percentage of notes generated in the ED.

Why Errors Occur
The errors identified in the study are likely caused by one of two possibilities, Weiner believes. "The first is limitations of the technology, which is rapidly improving but isn't perfect," he says. Users must speak clearly and use words the software understands.

The second reason centers on the ED itself, which has many unique aspects that Weiner says can make accurate and careful dictation more challenging. "The environment is busy, and documentation usually happens in a fragmented fashion," he explains. Physicians may dictate a patient's history and then be interrupted by a phone call, or another patient might need to be seen before the documentation can be finished. "It's just a very chaotic environment for doing dictation," Weiner says.

Despite being notoriously noisy and hectic, the ED may be an ideal environment for studying speech recognition errors. It's an arena that makes on-the-spot proofreading and editing difficult, giving early errors that may have been caught and corrected in another setting a greater opportunity to persist in the medical record. However, adopting best practices in the use of the technology and its hardware may give clinicians a leg up on ensuring better accuracy. "By using handheld microphones and only recording when necessary, the impact of any background conversations can be reduced," Ruff says.

Workflows are also likely to impact error rates when using speech recognition software in the ED. Kraus says the editing function must be built into the process in order for mistakes to be identified and corrected. "In an ED, they're typically busy, so is anyone taking the time to proof that information prior to submitting it into the record?" she asks.

Errors may also find their way into the record—and remain there—because providers often look at a limited subset of information in the EHR, which can lead to errors being overlooked. "Physicians might basically go to the end result," Kraus explains. "Is the person being admitted or not? And what is the diagnosis?"

If mistakes exist in portions of the record that aren't scrutinized by subsequent viewers, finding them postdocumentation may be difficult, she says.

Strategies to Avoid, Correct Errors
ED physicians can limit speech recognition error rates through diligence and a commitment to accuracy. Knowing how to speak to the computer is the first step, according to Weiner. "When talking, you have to be sure you're enunciating, that you're not mumbling," he says. Being aware of medication brand names and their potential to confuse a speech recognition system is also important. "If a new medication comes up and you use the name, it might not be in the dictation software dictionary yet," he says. Proofreading is the final and possibly most important step in reducing error rates. "No one wants to do it because it takes a lot of time but, based on our findings, we strongly recommend it," Weiner says.

Efforts to boost dictation competency should focus on improving patient care, especially in an era when the industry is emphasizing the medical record as a vital information-sharing tool. "[The medical record] is used for communication with other clinicians who are looking after that patient," Weiner says. "If that communication breaks down, it could lead to patient safety issues." A review of potential medical liability issues may also be useful in keeping physicians on track with their clinical documentation quality efforts, but "that shouldn't be the majority driver," Weiner says.

Downstream Opportunities
The study examined notes retrospectively, meaning any errors persisted—years later, in this case—in the medical record. They made it past any proofreading efforts as well as reviews that may have occurred by other providers involved in the patient's care. "Documentation errors may be found at any point and by any number of people who have access to a medical record, but there must be a system in place to correct—especially those errors that have the potential to impact patient care," says Guston, who points out that many facilities have document integrity programs, where analysts review clinician-created documentation, for just this purpose.

Even with many tasks now being automated, Zook says human oversight is still necessary. "There has to be either the physician or a back-end transcription editor looking over the report to correct errors prior to it being admitted into the EHR system," she says.

One option is to have physicians self-edit, although ensuring they remain committed to the effort can be difficult. "It comes down to them being responsible for that report all the way through," Zook says, noting that a concurrent review process can nab errors while the patient is in-house, and HIM may be called on to detect and correct mistakes in a postdischarge scenario. Either way, she says time and resources must be dedicated to the task.

To correct mistakes once they've entered the EHR, a resolution process must be in place. "Can the coder submit a correction? Does the correction have to be done by the original author? A lot of that depends on the institution's policy," Kraus says.

Rather than waiting for errors to persist further, clinical documentation improvement specialists may be able to connect with the physician soon after the record is entered to facilitate revisions. "Then they could make the correction right then and there," Kraus says.

Speech Recognition in the ED on the Rise
The fast-paced environment could make speech recognition a front-line tool for ED physicians, Weiner says. "If you think about a typical shift, you probably see 20 to 40 patients," he says, pointing out that documentation requirements continue to grow. "If you had to sit down and type everything out, it would be nearly impossible to do."

Speech recognition software has provided Brigham and Women's Hospital ED physicians with a weapon to keep pace. "It has been both a time saver and has increased the quality of work because it's much easier to talk and have your words recorded as opposed to typing everything," Weiner says.

Documentation emanating from the ED requires a short turnaround from the time a patient arrives and its availability to staff. "[The ED] is a natural fit for the speed that speech recognition provides," Ruff says. The technology may also serve as a key productivity tool that allows busy physicians to "avoid the alternative of not having documentation in the EHR at all or not having documentation in time," he adds.

As platforms improve their optimizations targeted at specific medical specialties—emergency medicine among them—their adoption in these challenging environments is likely to increase.

Adoption Pros and Cons
Guston says one of the most significant benefits of deploying speech recognition in the ED is the immediate turnaround time. "Documents are completed in real time, so if you have patients admitted through the emergency department, their documentation is immediately available to the care team who may need to reference them," she says.

Any cost savings associated with the adoption of speech recognition in the ED must not come at the expense of having errors slip through because there's no time to ensure the information is accurate, Guston says. "Unidentified errors have the potential to perpetuate throughout the medical record," she says, adding that this could potentially lead to safety concerns if care decisions are based on erroneous information.

Nevertheless, the speed at which speech recognition technology makes documentation available across multiple caregiver teams within the hospital stands to be a huge benefit. Staff rotation schedules and the need for efficient patient hand-offs in a short period of time are factors EDs may consider when exploring speech recognition solutions. "There could be an overflow of patients in the ED, with several physicians instead of just one," Zook says. "That may not be common, but it still impacts care. Documentation plays a huge role in easing that continuity of care and making sure each patient is taken care of in a timely manner."

With more EDs coming under the management of an outside physician group or other entity, Kraus says interoperability and training could be issues as speech recognition is introduced into these environments. "One of the issues with implementing speech recognition in the ED is that you have a lot of different people," she explains.

With little control over which physician is working during a shift or how their hours are scheduled, it's likely the hospital may see numerous care providers rotate through the ED. "The more people you have working in the system, the more potential you're going to have for errors because they're not necessarily trained on your system," Kraus says. Ensuring each physician is well versed in the use of the hospital's particular EHR platform as well as its speech recognition software could be a challenge. Without sufficient training, the organization may run a greater risk of errors.

Besides the risk of documentation errors—a concern that exists in nearly every department—Weiner doesn't see any downsides to using speech recognition in the ED. "It allows you to do a much more complete note. In the time it takes to type a brief sentence, you can quickly dictate a paragraph that captures your complete medical decision making and the history of that patient," he says.

— Julie Knudson is a freelance business writer based in Seattle.