Speech Recognition

 Well said, Mr. Radiologist. In a recent interview with AuntMinnie Radiology magazine, Dr. Giles Boland, medical director of teleradiology and vice chairman of radiology at Massachusetts General Hospital (MGH) in Boston, talks about the pressure Radiologists are facing when it comes to documentation. 

The article starts off with a rather capitalistic - although realisitic- view of the 21st century reading room: “radiologists today are measured constantly, whether it is in the number of images they read, their relative value unit (RVU) activity, or their report turnaround times.” Not only that. ”You can get three different radiologists to look at the same scan and they can give very different lengths and styles of interpretations,” Boland continues. “How is a referring physician to navigate through those variable styles?”

The advantage of voice recognition is that it may be counterintuitive because if radiologists have to edit that report, they naturally will shorten the number of words they put in it. They don’t want to edit a report that is five pages long.

According to Boland and colleagues as per another interview with European Radiology (March 8, 2008), the adoption of an “integrated RIS/PACS and voice recognition system” is the only path to ”reducing report turnaround consistently meeting stakeholder expectations.” Voice recognition systems “also offer the opportunity to create standardized, higher-quality reports,” they wrote.

Aunt Minnie editor goes on to comment how “MGH has utilized voice recognition technology for the past 11 years. During that time, the facility has reduced the average time it takes staff to go from a preliminary report to final sign-off to approximately three hours. Total report turnaround — the time from when a study is completed to when the final report is available on the system — is 12 hours.” Boland confirms:

“The majority of the reports that are signed in this institution are within a matter of a few hours of the exam having been completed. The heightened expectations have not increased reading errors by radiologists. Instead of leaving work unfinished, radiologists also are more inclined to complete their assignments.

“There is the (amount of time between the) completion of the exam to preliminary report, which really hasn’t changed much. That is a matter of an hour or two from when it is read to the time the exam is completed. The big change is the time from the preliminary report to final signing. For those groups with no residents or fellows, preliminary reports go away, so turnaround is even quicker.”

“Radiologists have to understand what their work product is — and that is ultimately the radiology report. The clear enhancement of value in that report is when you add voice recognition, because you can get it out quicker and remove that whole redundant route with transcriptionists. You also can structure it with macros and templates, shorten the report, make it more succinct, and the whole report turnaround time is faster,” concludes Boland.

The borders of the professional speech recognition community are expanding further South with a new member as of today:

Drs Conidaris and Partners, a private radiology group located at the Glynnwood Hospital premises in Benoni, South Africa, have completed the rollout of a Crescendo/SpeechMagic speech recognition system. After evaluating different technologies, the partners chose SpeechMagic for its wealth of built-in Radiology vocabulary and Crescendo’s unique voice streaming technology, which allows for immediate speech processing; a key feature for mission-critical environments” explains Kevin McEvoy, Managing Director, Datafer.

With the Crescendo/SpeechMagic technology, the practice was able to deliver on its initial objectives: speed up report production and reduce transcription costs. With the secretaries’ newfound ability to edit medical reports as opposed to typing them in full, fewer resources are required to process the same volume of reports. The first medical report dictated achieved over 75% accuracy and this rate is continually improving as the system learns.

“Skilled medical typists are extremely hard to find in the region, and a growing number of South African healthcare facilities welcome speech recognition technology as a reliable, cost-efficient way to address this issue. The technology has clearly matured significantly over the past few years and is delivering the expected results,” explain the radiologists at the practice.

Digital mobile recorders (Philips 9600) are used by physicians to dictate either directly from the practice or on the move. Every time the device is docked and a network connection is established, DigiService-IP automatically and securely streams completed voice files to the central server in real-time. The Secretary then uses DigiPlayer-IP, the Microsoft Word based playback and transcription application from Crescendo, to correct the document.

“South African hospitals and clinics are increasingly aware of the importance of digital technology to modernize both their practice and care delivery, and they are very selective in their purchasing process, with good reason” explains Costa Mandilaras, President, Crescendo Systems Corp. “This is why we chose a distributor, Datafer, with outstanding support services to properly deliver the Crescendo 18-year field expertise to South African customers. I believe that Drs Conidaris and Partners is the first of many more South African healthcare installations to come,” concludes Mandilaras.

The Best-Practice Fairy is back with a fascinating article by Jeff Kelly on successful speech recognition implementation. No doubt that Mr. Kelly knows what he is talking about: before joining Inland Imaging, a 60-member private radiology group in Spokane, WA, as director of clinical applications, he used to perform RIS market analysis and consulting work for MedQuist on voice recognition implementation. As Kelly puts it in a 2007 AuntMinnie article, one question - “What’s in it for me?” - is the key to a speech recognition project’s success. His article sheds the most promising light on the implementation brain-teaser. Here are a few soothing extracts:

Strong support and leadership

Consider pulling in this diverse group of representatives into a VR task force that can meet at defined intervals to address issues and help break down barriers to success. Then use this task force to educate its members on voice recognition — keeping in mind that this is usually new technology for 90% of those involved. This task force can also function as the decision-making body behind the implementation.

Maximize benefit, minimize effort

A typical radiologist statement that must be addressed is, “I am not a transcriptionist.” You can respond to this statement on three levels. First, stress the improvement in patient care when the radiologists are able to review and edit the report at the same time as reviewing the films. No longer do they need to recall whether it was a left or a right shoulder MRI. The accuracy of the report immediately increases.

Second, by editing at the time of the report, they eliminate the time-consuming process of logging into the RIS/HIS at a later time and reviewing the report. Third, the ability to send to transcription still exists if they so desire.

Provide sufficient internal support

At Inland Imaging, an in-house trainer was designated to be with each radiologist for the first three days when they began using the technology. This level of support tapered to half of every day for the next three days, then reached a point in which only a phone call was necessary to check in on each radiologist. Keep in mind that each radiologist will be unique — some needing more support and some needing less.

Set realistic expectations

As stated before, the radiologists are going from a traditional bicycle they’ve ridden their entire life to a shiny new two-seater that they are going to share with a friend. This is no easy task. It’s important to be frank and honest with staff and radiologists about the implementation process. There will be application bugs. There will be times when the application or workstation stops responding. There will be days when they love the system, and days when they feel like throwing the whole thing out the window. Set these expectations up front, then address them as quickly as possible when they arise.

Making change last

Finally, maintain the health of the system. Continue to refresh hardware as necessary to optimize both client- and server-side performance. Maintain error logs and database optimization jobs. Continue to accept version upgrades. Don’t allow neglect of system maintenance responsibilities to have a damaging impact upon the workflow of its users.

> Read full article (registration required - free)

> More threads like this

Someone asked me to provide a few examples of healthcare facilities having deployed SpeechMagic in the front-end (process whereby the recognized text appears on screen as the physician dictates). It is my pleasure to oblige:

• Alegent Health, Omaha, Nebraska, USA
• Staten Island University Hospital, Staten Island, USA
• Jewish General Hospital, Montreal, Canada
• Aberdeen Royal Infirmary (ARI), Aberdeen, UK
• Ulleval University Hospital, Oslo, Norway
• Bornholm Hospital, Bornholm, Denmark
• Helsinki University Central Hospital, Helsinki, Finland

For more information on front-end/back-end speech recognition workflows, see this thread.

With an industrial history firmly rooted in technology, the Nordics have always been early hi-tech adopters, clearly ahead of new market trends. While, in the early 2000’s, the rest of Europe was still getting used to the very concept of online transactions, these guys were already buying books by the dozen from Amazon.com over broadband connections. The fact that Norvegian healthcare is about to roll-out the world’s largest speech recognition installation therefore comes as no surprise.

As announced yesterday, over 1,000 physicians from the Ulleval University Hospital in Oslo - Norway’s largest clinical center - will be using SpeechMagic to generate reports across all medical specialties as of 2009. This is expected to be the world’s largest deployment of front-end speech recognition at one single hospital site. > More…

On the other side of the northern border, their Finnish neighbours also decided to step aboard the speech recognition ship with zero drowning reported so far. Instead, a rather impressive use case from the Palokka healthcare joint authority, who has been using a speech recognition system powered by SpeechMagic since December 2003. The system allowed the authority to “speed up patient care and improve data security at the same time. Patient data is always in the right place at the right time, which avoids unnecessary processing and reduces the need for doctors to consult each other.” > More

The word is clearly spreading. A recent CRM Buyer Magazine article reports positive results for two US hospitals that have adopted speech recognition technology for their respective ER Departments. The news echoes the 2006 HIMSS survey, in which 65% of chief information officers had announced plans to equip their staff with speech recognition technology by 2008.

Abbott Northwestern Hospital, Minneapolis
Speech Recognition has been around for the past two years in the hospital’s ER department.

I’m able to complete my charts and consult other physicians about patients in real time. In the past, you might not see dictated notes for six to 12 hours, but now it’s instantly accessible by the entire team,” comments Christopher Obetz, M.D., emergency medicine physician.

The technology is not without its problems though. As you’re dictating, you have to keep one eye on the computer and make sure its transcribing correctly. Compared to the traditional method, a lot of the burden falls on the user for accurate transcription.

Ironically, while the software does very well with complicated medical language, it often struggles with shorter words or phrases such as “I am” or “she is.”

Massachusetts General Hospital, Boston
After a successful trial in the ER and a larger pilot in 2005-2006, speech recognition is now used by approximately 175 providers in 30 departments throughout the hospital.

For the most part, the software is easy to integrate with relatively short set-up and training times,” says Deborah Adair, director of health information services.

Harry Rhodes, director of practice leadership for AHIMA, commented the above news:

It’s being touted as a natural add-on to the electronic medical record, since doctors are used to recording their notes. The software can also prompt users to include information they may have forgotten, creating a more complete record. Resistance to change is also a hurdle. It’s a big transition, but not an insurmountable one, given that the technology can improve quality of care.

See also: wireless speech recognition in the ER- case study

Implementing speech recognition for radiology reporting before a hospital mandates can be advantageous for radiology practices, Dr. John Floyd of Radiology Consultants of Iowa told attendees at an informatics scientific session the RSNA 2007 meeting. The session was reported by Editor Cynthia Keen in the AuntMinnie Radiology Magazine:

Radiology Consultants of Iowa in Cedar Rapids serves two urban hospitals, seven rural hospitals, and one imaging center. In 2005, with a 37% error rate for internally transcribed reports at one hospital, and a report turnaround time of three to 10 days at another, executives at both hospitals were discussing the purchase and adoption of speech recognition systems as a way to alleviate the problems.

A Single SR System

Floyd said that he and his colleagues recognized the inevitable, and offered to take over all aspects of radiology report transcription, processing, and distribution. After evaluating different systems for six months, the practice purchased a single speech recognition system to be interfaced with five different RIS and three different PACS.

The benefits of having a single system were that the radiologists and transcriptionists/editors would only need to learn how to use one system, rather than several, and would only need to work with one database with respect to adding vocabulary and teaching the system to correct errors.

The radiologists planned to phase in usage in the first hospital to go live over a several-month period. “However, the speech recognition system launched abruptly in March 2006 when the transcriptionists learned what was being planned and walked off the job,” Floyd said.

A Majority of Free-Text Reports

“We self-edit 97% of approximately 1,000 unstructured reports each day,” Floyd said. He believes that the practice’s ownership of the speech recognition system contributes to this high level of self-editing compliance.

Approximately 90% of the reports generated are free text, with less than 10% using standardized responses. Transcription editors are available from 6:00 a.m. to midnight, but are typically used only when a report is particularly complex or when data need to be entered with the report. “This can add an additional five to 10 minutes of time,” Floyd explained.

Comparing Error Rates

Concern about high error rates of SR systems reported at RSNA 2006 convinced the practice to conduct an analysis of accuracy of SR-generated reports as compared to reports still being generated through traditional dictation at one rural hospital. The group compared 493 SR reports with 283 traditionally transcribed reports. The total number of errors for the SR reports was nine, compared to 13 for the traditionally transcribed reports, and 0.6% of the SR reports had a significant error, compared to 2% of the traditionally transcribed reports.

Productivity Gains

Although a direct correlation cannot be attributed to the implementation of speech recognition, Dr. Floyd said that during the remaining months of calendar year 2006, productivity increased 12% as it related to reporting. In the first eight months of 2007, productivity increased by 28%. Floyd attributes financial ownership to making the project a success.

Integrating Dictation with PACS

In the RSNA same session, a group from Santa Barbara Cottage Hospital reported that by integrating a dictation system using structured templates with its PACS, a reduction of 45% in full-time employee (FTE) transcription was achieved within 90 days from implementation in February 2006. Compared to the same months of the prior year, report turnaround time for STAT exams improved by 70% and overall turnaround time improved by 8%.

> Link to AuntMinnie web site

As previously discussed on this blog, the ER Department of the Sir Mortimer B. Davis - Jewish General Hospital in Montreal, QC, introduced a wireless document creation and management system that uses PDA devices and speech recognition to accelerate the delivery of critical information at the point of care. Just released, the case study provides details on Dr. Rsoenthal’s vision behind the project, key project milestones and achievements.

> Read case study

Freshly inked article from Australian IT magazine reads: “Voice recognition in medico lingo”. Editor Karen Dearne interviews Chief radiologist K.C. Fan from the Westmead Children’s Hospital in Sydney West.

The artcicle starts with a bit of history on professional speech recognition, quoting Dr. Fan’s experience:

…the technology has taken a while to achieve the accuracy they require. Fan says early versions were inaccurate and too slow. The processing time for computers to turn voice into text reduced the potential for productivity gains.

Tailored dictation products began to surface a couple of years ago, and the most successful voice recognition applications have been in structured environments, in which vocabulary and context can be built into transcription engines.

The author goes on to describe the actual project in place at the Radiology Department of the Children’s Hospital, with a very accurate insight from Dr. Fan on “how the limitations of general speech systems are actually strengths in specialised applications, such as radiology reporting:”

In day-to-day usage, short words such as to could also be two or too. The software has to guess what you mean and try to put the correct word into the context of your sentence. That’s not a major problem in radiology because specialist medical terms are long and complex. That’s an advantage, as it’s easier for the speech recognition system to pick them up.

A word such as interceception (inverted gut) comprises five syllables, for example, providing a clear, machine-readable pattern. Conversely, Fan says, most people unfamiliar with the term would misspell it. That sort of accuracy is a challenge to humans, but not to the system software.

When it comes to the difficulty of finding skilled medical transcriptionists, the situation in Australia seems very similar to that in the USA:

A good medical typist will get it right most of the time, but it’s hard to find skilled people.

On the dictation side, Dr. Fan stressed the importance of a dictation etiquette:

It is important for users to be consistent in their dictation. If you change the way you speak, you can teach the system bad things and the accuracy will drop. If people are consistent, it doesn’t matter whether they have an Indian, Chinese or Irish accent, the system will adapt.

You can read the full article, with further details on RIS integration and benefits, directly from the Australian IT web site.

Hi! I’m the Best-Practice Fairy. My mission is to dig up articles that cover the implementation aspect of speech recognition technology within healthcare settings. Here is one.

In the October issue of Health Imaging & IT, Journalist Beth Walsh interviews a number of key North American healthcare executives about their experience deploying speech recognition technology. This article really caught my attention as I believe it offers what healthcare professionals really want: a way to learn from each other. Whether it is in a Radiology or ER setting, implementation tips converge: all respondents insist on the importance of a gradual rollout, as well as on a the importance of interfaces to make sure that the technology fits the way physicians work, and not the other way round. This is particularly key when it comes to using speech recognition in the ER, where mobility is the name of the game. Here are a few extracts:

Implement Voice for the Right Reasons

Implement voice for the right reasons, says Selle. “Our goal was not to make money off the system, but to provide a service to the facility and improve our marketability to the community and referring physicians.”

Emergency department notes must be relatively up-to-date, says Rosenthal. “In other areas, you can wait to see notes.” And since two-thirds of the hospital’s ED volume occurs outside of the standard 9-to-5 day shift, “voice recognition offers a good opportunity to do notes whenever we need them,” he says.

Project Set-up

Both Selle and Rosenthal say that the implementation of voice recognition warrants a thorough evaluation of workflow and an infrastructure geared to the technology. “Interfaces needed to be set up between facilities’ management systems to generate orders,” says Selle. That way, you’ll have the same patients on the voice worklist as are on the PACS or other systems clinicians are reading off of. He recommends setting up destinations for reports once they are dictated and physicians have signed off. “Set up network printers in the ED that will print reports directly to clinicians,” Selle says. “There’s also the interface on the backend to send the reports electronically into the facility’s systems so that they’re accessible and combined with patient information.”

Rosenthal says it’s common for hospitals to underestimate the setup required for voice recognition. “You need a vision of where [the technology] will go,” he says. For example, if your bandwidth is too slow, people won’t want to use the system. Plus, you won’t have the capability to eventually expand to nurses using voice recognition and performing electronic documentation, for example.

Overcoming Resistance

Rosenthal says that new technology will only be successful if it adapts to the way people already function. “You can buy the nicest solutions, but if people have to go out of their way to work in a different manner, they won’t use it.” That principle has played out, he says, as vendors are offering more ways to use voice recognition and bringing it into various applications. “The technology is fitting more into the way people work.” High failure rates in the early days of voice recognition were due to systems that asked busy clinicians to do things they wouldn’t normally do, he says.

Aurora Health Care, a 14-hospital, not-for-profit system covering eastern Wisconsin (…) purposely implemented voice recognition on a gradual basis. One reason for that was physician resistance, Hartig says. “I think resistance is normal. Most people resist change. We wanted the end-users to embrace it. We didn’t want to force it on them.” The organization still had its old system from Dictaphone so the physicians could opt to use that system or the new PowerScribe solution. The doctors who did use PowerScribe were able to reduce their report turn-around times very quickly, Hartig reports.

Gradual Implementation

Rosenthal says it’s very important to implement voice in phases. “Don’t deploy [it] on a large basis until you’ve run a good pilot and it’s functioning well,” he advises. His facility has been using voice for follow-up notes and is in the process of progressing to more complex documentation. Rosenthal has noticed that people are more willing to share their experience and knowledge about technology than they were in the early days of voice. “What we’ve done in the ED is now being deployed elsewhere in the hospital. Why reinvent the wheel each time? If somebody else can benefit, I think that’s good.”

> Read full article