Voice recognition software has been shown to reduce report turnaround time and holds promise for populating and mining structured reports — but not all radiologists are convinced.
Many users still find the software cumbersome and error prone, as seen in a recent informal Diagnostic Imaging poll where 80 percent of respondents said they use it, but 30 percent of them reported frustration with the software. http://www.diagnosticimaging.com/voice-recognition/content/article/11361...
And radiologists have steep demands for the software, experts said. With their need for accuracy and use of complex terms, fast-talking radiologists really put the systems to the test.
“Radiologists are trying to provide an enormous amount of documentation in a very short period of time, accurately, and they are experts at grinding through that workflow and process as fast as they can,” said Joe Petro, senior vice president of research and development for Nuance Healthcare. “They hold us to a much, much higher standard.”
Over the years, the mathematics behind the software hasn’t changed much, Petro explained, but the accuracy has improved from 75 percent 10 or 15 years ago to the “high 90s for radiology” today. His company, as with many of his competitors, has been chipping away at the error rate by fine tuning the speech recognition algorithms and taking advantage of the computer chip power.
Jacques Gilbert, global marketing manager for radiology for GE Healthcare, said that the technology has gotten better at differentiating natural pauses in dictation and accents — and the GE tool learns from its mistakes. Their algorithm takes into account corrections a user makes, applying the change going forward.
“It learns upon itself,” he said. “In theory as the radiologist continues to a make change to the reports they are dictating, it’s getting better and better.”
That’s the main feature where the products compete, he said, in the time it takes to make the fixes. They also are competing based on how good the product is right out of the box, taking into account training time and accuracy, said Tim Kearns, GE Healthcare’s product manager.
|Check out the experts' tips for improving your voice recognition software experience.|
The training time might be just the snag that hung up many early adopters, experts said, accounting for some of the lingering frustrations with the software. Some users who still find it cumbersome might not have spent the time to understand the nuances of the experience, Petro said. For example, if the software is making the same mistake over and over, there are features that allow the user to modify it to avoid the mistake in the future.
“You can change your experience and you have to be committed to it in the short term,” he said. “The most successful physicians have spent the time and make that commitment, but the technology is evolving so that you don’t have to be that committed.”
Some radiologists need to be convinced of the software’s benefits, said Don Fallati, senior vice president of product management for M*Modal, which was recently acquired by MedQuist. It also has to fit into their workflow seamlessly and provide data mining tools to make the effort worthwhile.
“We still do encounter some areas of resistance or continued inquiry or challenge,” he said, adding they are “proving or demonstrating that it is realistic and productive and that it meets the usability test.”
Indeed, most radiologists embrace the technology, Petro said, because it can make their lives easier. And as the software continues to evolve, it will provide opportunities in how the vast amount of data dictated is mined and shared.
“It’s beyond speech,” Petro said of the future of voice recognition.
Through what’s known as natural language processing, vendors are exploring how to develop systems that can really understand what the radiologist is saying, and populate health IT systems accordingly. Natural language processing takes the unorganized narrative, codes and structures it, and harvests it for specific data. The idea is that items such as a patient’s problem list can be automatically sent to the electronic health record, or certain decisions can be automatically checked against appropriateness criteria.
“We want to move beyond just documenting the patient information [to] populating repositories and databases,” said Fallati. “We need to get to structured data. They want intelligence, not just data.”
Speech recognition will also become more tightly integrated into RIS and PACS, as well as EHRs, Fallati said. Information will automatically flow into these systems from the dictation for robust documentation and feedback.
It’s an approach being taken by Intelerad, which imbeds voice recognition into their software packages, saving on real estate, as Sebastien Cadet, product specialist leading Intelerad’s VR projects, explained.
“It’s a seamless process for the end users and they literally deal with one application for all their work flow and one vendor for the service and support of that application,” he said.
Similarly, future systems will be automatically populated in templates with patient data, for example, saving the radiologist time and effort in the dictation, said Gilbert. “It’s a bit of a frustration for the radiologist to have to do that when they can see the data is in the system,” he said. “There’s a tremendous amount of time savings when the data they’d like to see in the report from the RIS or PACS are in the report in a structure that they have decided on for different procedure types.”