Speech systems earn respect as technology boosts productivity

Despite the ability to save radiology departments hundreds of thousands of dollars and dramatically decrease report turnaround time, speech recognition (SR) systems have long been denied market acceptance, perhaps because those savings have come at the expense of radiologists. Within the past two years, however, healthcare economics and advances in speech system technology have converged.

Productivity is winning out. The routine use of speech systems at about 300 sites in the U.S. has led to investment returns that are not only impressive but documented.

“Speech recognition affects two metrics by which radiology departments are measured-cost and turnaround time,” said Michael Mardini, president of Talk Technology of Bethsalem, PA. “From a turnaround perspective, it’s a no-brainer. The radiologist looks at the images, dictates and signs the report, and it’s out the door.”

In the past year, speech recognition vendors have reported an upswing of interest by prospective customers, primarily radiology administrators. That interest has prompted partnerships with PACS vendors to distribute workstations equipped with both technologies.

A key indicator that SR has come of age was the April acquisition of Talk Technologies by Agfa (SCAN 4/25/01). In addition, most of the major PACS and RIS vendors now offer an integrated system that incorporates SR applications, with Talk Technology and Lernout & Hauspie products at the forefront of the trend. Talk Technology has its SR system installed at 200 sites.

Adding speech recognition to the request for proposal accelerates the anticipated return on investment for PACS by a year, according to Mardini, who calls the relationship symbiotic.

“The major trend today is integrating speech solutions into RIS and PACS,” said John Wald, vice president of sales at L&H. “A PACS installation that relies on traditional transcription and turnaround times defeats the goal of increased efficiency.”

Few would argue that the traditional dictate-and-transcribe mode of radiology reporting has problems. Conventional transcription, whether in-house or outsourced, is costly, and preliminary reports are not available promptly. Even when fully staffed, transcription services may not be able to cope with the peak volume of dictated reports, causing delays of hours or days.

Solving those problems with SR requires a technological leap on the part of radiologists from Dictaphones to computer software. It also demands a change in the work process-shifting report editing from transcriptionists to physicians. Even with today’s advanced SR systems, physicians still spend more time dictating reports using SR than they did relying on traditional transcription. The cost savings reaped by the institution benefit radiologists only indirectly, if at all.

“No one disputes the fact that it saves money,” said Dr. David Weiss, chair of radiology at Chestnut Hill Hospital in Philadelphia. “The controversy is, whose money is it saving? Radiologists are angry because they’re being asked to do extra tasks to save the system money.”

For many departments, however, SR is the key to making the most of limited resources while staying competitive. Three years ago, Weiss began tracking the concerns cited by referring physicians about the services provided by Chestnut Hill’s four radiologists. These included delays in receiving radiology reports and problems with accessing patient films from among the 400 exams conducted in the department daily.

At Chestnut Hill, the radiology team went cold turkey and dropped use of traditional transcription within a week of SR implementation. Return on investment was achieved in about one year, Weiss said, and the department is now realizing savings of about $100,000 annually.

The current generation of SR systems represents a substantial improvement over earlier technology, which even five years ago was not as good as it is today, Mardini said.

“Four years ago, the market was nothing. Maybe 10 to 15 sites were using speech recognition to some extent,” he said.

The biggest advance since then is the move from discrete to continuous speech systems, according to Art Augustine, project engineer for healthcare devices at ECRI in Plymouth Meeting, PA. The old discrete systems required measured speech with pauses between words for best recognition accuracy. Continuous systems are capable of recognizing natural speech. In addition, advances in computation power have enhanced the ability to support an increasingly complex language model, which not only recognizes natural speech but is also capable of anticipating speech context.

Accuracy and the time radiologists spend correcting errors remain the biggest stumbling blocks for SR, but prospects are looking up. Vendors report accuracy rates averaging between 97% and 98%, an optimistic accounting but one that has landed in the ballpark, according to Augustine, who is conducting a performance evaluation of commercially available systems. He reports that average accuracy hovers around 90%, and anecdotal evidence supports this assessment. Long-time users with a high degree of familiarity and proficiency with SR say that 90% accuracy rates are about average in day-to-day use.

Although acceptance has increased dramatically in the past year, speech technology is still far from being enthusiastically embraced by radiologists, Mardini said. At any given site, the majority of radiologists will accept SR, albeit grudgingly; 10% will hate it; another 10% will embrace it.

“I think no matter what we do technologically, we will always have about 10% of physicians who hate it,” Mardini said. “There will always be a percentage of radiologists who don’t want to change.”