The effect of voice recognition software on comparative error rates in radiology reports

This study sought to confirm whether reports generated in a department of radiology contain more errors if generated using voice recognition (VR) software than if traditional dictation-transcription (DT) is used. All radiology reports generated over a 1-week period in a British teaching hospital were assessed. The presence of errors and their impact on the report were assessed. Data collected included the type of report, site of dictation, the experience of the operator, and whether English was the first language of the operator. 1887 reports were reviewed. 1160 (61.5%) were dictated using VR and 727 reports (38.5%) were generated by DT. 71 errors (3.8% of all reports) were identified. 56 errors were made using VR (4.8% of VR reports), whereas 15 errors were identified in DT reports (2.1% of transcribed reports). The difference in report errors between these two dictation methods was statistically significant (p = 0.002). Of the 71 reports containing errors, 37 (52.1%) had errors that affecting understanding. Other factors were also identified that significantly increased the likelihood of errors in a VR-generated report, such as working in a busy inpatient environment (p<0.001) and having a language other than English as a first language (p = 0.034). Operator grade was not significantly associated with increased errors. In conclusion, using VR significantly increases the number of reports containing errors. Errors using VR are significantly more likely to occur in noisy areas with a high workload and are more likely to be made by radiologists for whom English is not their first language.