Speech recognition technology finds a voice

By: Merlina Trevino

It's been one of the most popular Diagnostic Imaging PACSpoll questions: Are you ready to replace traditional transcription services with speech recognition technology? More than 50% of the poll respondents answered with a resounding yes. The buzz surrounding speech recognition promises better accuracy and faster report turnaround, ultimately enhancing patient care.

But overworked radiologists won't jump onboard without a lot to say about how, if, and when speech recognition should be implemented. Presenters at the 2002 Symposium for Computer Applications in Radiology in Cleveland looked at real-world case studies and discussed technical issues, financial advice, and more than a few caveats about speech recognition implementation.

Some of the most basic questions around speech recognition technology involve how to implement it. Should it be a stepwise process, slowly phasing out the transcriptionist? Or should implementation be done in one fell swoop, ripping the transcription pool out of the equation quickly and making the transition as short and painless as possible?

"Basically, there are two ways to implement speech recognition technology," said Alan Schweitzer, manager of technology services at the Radiology Consulting Group in Boston. "The first is minimally invasive for the radiologist; it involves implementing the technology but still using the transcriptionist as an editor. The problem with this method is that you don't get much benefit in report turnaround time because the transcriptionist is still in the loop. By keeping the transcriptionist on board, potential cost savings are limited to approximately 50% of current costs."

Other institutions have skipped any transition period and mandated the complete rollout of speech recognition technology all at once.

"Just before I came here, the radiology department's transcription situation was desperate," said Dr. Chris Sistrom, an assistant professor of radiology at the University of Florida. "Report turnaround times were sometimes a week or more, and the trend was worsening. Out of necessity, the department just jumped in and went cold turkey into speech recognition. The system was installed over a weekend, and the following Monday all reporting was done with speech recognition. To say the least, it was a culture shock, but they did it."

MONEY TALKS

Perhaps the biggest incentive for installing a speech recognition system is the potential for large financial savings that automation always promises. Installing a speech recognition system is supposed to eliminate transcriptionists' salaries, benefits, annual increases, and incentive pay.

Reporting on several case studies, Schweitzer highlighted additional "soft" benefits not readily apparent, such as a reduction in clinician, radiologist, and support staff time spent tracking down diagnostic reports. He cited the availability of more timely information to support clinical decisions, resulting in improved patient care, lower healthcare costs, and reduced length-of-stay for inpatients. The enhanced level of service to referring physicians translates into a competitive advantage and potential growth in revenues.

Massachusetts General Hospital, for example, performs 550,000 radiological exams a year and produces 1500 diagnostic reports per day. The 800-bed teaching hospital first began its speech recognition deployment in September 1996. Since the adoption of speech recognition, MGH has shown a total transcription cost reduction of $629,881 from 1997 to 2001.

Similarly, at Children's Hospital Boston, which performs 140,000 radiological exams per year, the transcription savings jumped from $73,500 in year one to $212,157 in year five. Cumulative cash flow rose from $226,677 in year one to $362,793 in year five.

INCREASED PRODUCTIVITY?

Some radiology departments actually reported a decrease in productivity in the first few months after the initial implementation of speech recognition-not a strong argument in favor of its adoption. This decreased productivity could very well reflect the way a department implemented the technology rather than the efficacy of the technology itself, however, said Steve Langer, Ph.D., senior associate consultant of diagnostic imaging at the Mayo Clinic in Rochester, MN.

He described a study performed at the University of Washington, Seattle that outlined four different methodologies for implementing speech recognition into a radiology department: film with human transcription, filmless with human transcription, film with speech recognition, and filmless with speech recognition. The first scenario involves seven steps from patient arrival to hard-copy report. The fourth scenario reduces the number of steps to three from patient arrival to final uploading of the report to the RIS, PACS, and electronic medical record. Reducing the number of steps decreases report turnaround time by streamlining the human interfaces involved.

Surveys in the ongoing study were sent out to more than 40 North American sites and have been completed by 11 of them so far. The surveys were developed to measure the impact of speech recognition technology on report turnaround time and report productivity. Productivity was defined as the number of reports per day divided by the total number of full-time radiologists.

Speech recognition adoption cut down report turnaround time for institutions still using film by 21%, and productivity more than doubled over the film/transcription methodology. Institutions that adopted PACS without speech recognition were able to cut report turnaround time by 40%, while report productivity remained the same. Langer attributes this paradox to the fact that radiologists who spent less time waiting for cases may have spent more time reading them on PACS than on film.

The final scenario of integrating PACS with speech recognition produced the best results, reducing report turnaround time by more than 60%. The sites implementing this model did not, however, exceed the report productivity of sites using film and speech recognition. Langer explained that 16% of these sites had not fully adopted speech recognition, and in some cases they used transcription services as a backup. As departments become more familiar with the technology, that productivity should increase, he said.

BARRIERS TO SUCCESS

Despite promises of increased cost savings and decreased report turnaround time, many radiology departments remain hesitant about moving to speech recognition. The Diagnostic Imaging PACSpoll found that nearly 50% of respondents were either unready or unsure about making the move away from transcription. Mistakes made during the initial rollout led to productivity horror stories that don't do much to sell the next round of radiology departments on the technology. One concern in the past had to do with the accuracy of the technology itself (see accompanying story), which may have been the case during the implementation of the first generation of speech recognition. But some radiologists note that the technology today is accurate enough to provide a viable replacement to transcription.

"The biggest misconception radiology departments have is that the technology is not ready and if they wait longer it will improve," Schweitzer said. "But this is like contemplating the purchase of a home PC and being afraid to commit, knowing that in six months something better will be available. The stability and recognition accuracy of the technology is improving over time, but at some point you simply have to take the leap. The report throughput and cost-savings benefits are realizable now, and many institutions have demonstrated that the technology can work."

Another misconception is the idea that speech recognition will increase productivity right out of the box. As with any new technology, a fair amount of training must take place before all of the kinks are worked out.

"Too often radiologists are not fully aware of the learning curve that will have to be climbed to be truly efficient with current products," Langer said.

It's imperative for a radiology department to provide constant technical support, or what one of Schweitzer's colleagues calls "hover training."

"Speech recognition technology is deceptively simple to learn," he said. "But in reality, you need a couple of weeks, if not longer, to completely learn the technology. What's really needed-and difficult to find-is someone in technical support who has the necessary technical savvy but is also a good trainer."

Common mistakes made during initial implementation of the technology run the gamut from understaffing and underfunding the initial rollout to something as simple as adjusting the input volume of the microphone, which can be critical to improved accuracy.

"The most common mistake radiologists make is getting hooked on the technology itself," Sistrom said. "Total reporting time is increased and film viewing time is decreased when radiologists watch their words appear on the screen rather than look at the films. The second most common mistake is forgetting the way we used to work and failing to consider using batch mode for high-volume work."

Some radiologists claim that they can't work in batch mode because they may forget what they saw or what they were describing in a particular exam if they hold off until after dictating a batch of reports before editing them, Sistrom said. Yet this was the way radiologists used to work with plain film, waiting to edit reports once they came back from the transcriptionists. Sistrom said that it's just a matter of realizing that working with speech recognition in batch mode actually turns reports around faster than in film days.

TEMPLATES AND MACROS

Even with training, radiologists may find speech recognition too slow. Sistrom outlined how templates and macros can go a long way to increasing productivity. He defined templates as representing a predefined, complete radiology report, including text and dynamic editing elements, so that a radiologist can customize it to the case at hand. A macro is a predefined text fragment that may be as small as one word and as large as a paragraph.

"Learning to use templates and macros did regain some productivity lost during the transition to speech recognition," he said. "Making and using templates and macros gives radiologists a feeling of participation and control rather than simply being frustrated with the new technology."

When using templates, radiologists should consider what type of exam they are dealing with first, according to Sistrom. High-volume, short, and simple dictations, such as ICU chest films, probably won't be helped by a template, which might slow the radiologist down.

"Fully realized structured templates are much more appropriate when reading a complicated case such as an MRI of the brain or CT scan of the abdomen. For these examinations, using a complete template is very helpful," Sistrom said.

The extra time a radiologist spends working with the template reporting interface while looking at the images can be made up by the decreased amount of spoken text needed. These templates can contain headings for each organ or structural/functional domain, such as a generic template for abdominal CT (see example opposite).

"In our dictation system, the bracketed portions of templates allow the radiologist to move from one to the other with a button on the microphone," Sistrom said. "The brackets and contained text are automatically replaced by what you say, making it fairly easy to navigate through and complete the report. I find that referring clinicians like this because all reports for a certain examination have the same order and format."

With the use of macros and templates in speech recognition technology, radiology departments can begin to move toward fully structured reporting. Sistrom envisions a complete radiology ordering, scheduling, procedural, and reporting system. He sees an integrated PACS and speech recognition process that would allow instant access to pertinent patient information from the minute the patient walks into the referring physician's office. Referring physicians would be able to enter a request for an exam and be automatically connected to that patient's online medical record.

After selecting a diagnostic indication from a menu or search facility, the doctor would then view a menu of possible exams with the most appropriate clearly indicated. When the appropriate examination is chosen, the patient would be scheduled for it automatically.

When the patient arrives in the radiology department for the study, a previously defined examination protocol would be automatically loaded on the imaging device and that examination done. When the radiologist selects the case, the images will come up along with a previously defined template for that examination/indication combination, providing a checklist for that interpretation and ensuring a complete, standard report for every case. After signing, the report would then go back into the online record and would be available for immediate review by a referring physician. In a critical situation, the system may page or e-mail the physician with a text message.

"In implementing this vision, more important than technological innovation will be groups of radiologists and clinicians working together and specifying appropriate examinations, indicating specific protocols and reporting templates," Sistrom said.

The key message about the implementation of speech recognition technology was clear: Success is dependent on the needs of radiologists as well as of the institution. Radiology departments should ask the right questions about what they need before they jump into speech recognition adoption.

"Is there a pressing reason to improve report turnaround times? Are human transcription costs out of control? Are the radiologists realistic about the demands that the new system will place on them and willing to proceed? Is there credible onsite expertise to keep the mission-critical system up? If the answers to all those questions are yes, then maybe the site should consider speech recognition," Langer said.