Automatic Speech Recognition Has High Error Rates in Breast Imaging Reports

The study reviewed 615 reports of complex cases discussed on multidisciplinary team rounds: 308 reports generated with automatic speech recognition and 307 reports generated with traditional dictation transcription. “Our study found that at least one major error was found in 23% of automatic speech recognition [ASR] reports compared to 4% of conventional dictation transcription report,” said Anabel Scaranelo, MD, from the University Health Network (Toronto, Canada). Key errors were described as those errors that had an impact on the understanding of the report and errors affecting patient management, such as an incorrect unit of measurement (for example, millimeter instead of centimeter) or a missing or added “no.”

The error rate was even higher if breast magnetic resonance imaging (MRI) reports were looked at separately, reported Dr. Scaranelo. The major error rate was 35% for ASR reports and 7% for conventional reports. “We think this is because MRI reports are more complex, with more description,” Dr. Scaranelo said.

Native language had no effect on the ASR report error rate, according to Dr. Scaranelo. “We thought that there may be a higher error rate for nonnative English speakers because the software works with voice recognition, but that didn’t happen,” she said. The ASR system is fed with several hours of dictation from the radiologists, and it appears to have “learned” from the voice data that was input, she reported.

ASR is typically used in several countries, according to Dr. Scaranelo. “The results of our study emphasize the need for careful editing of reports generated with ASR. They also show a strong need for standardized templates and use of structured reports, especially for breast MRI,” she said. The study’s findings were published in the October 2011 issue of the American Journal of Roentgenology.

Related Links:
University Health Network