Errors Common When Speech Recognition Software Is Used to Document Patient Visits

By Amy Orciari Herman

Edited by Susan Sadoughi, MD, and André Sofair, MD, MPH

Speech recognition software, used increasingly to document clinical encounters in electronic health records, yields errors in 7% of dictated words, suggests a study in JAMA Network Open.

The analysis included roughly 220 documents (e.g., office notes) that were dictated using one speech recognition system at two U.S. healthcare organizations. Researchers transcribed each original recording to use as the reference standard, and then examined errors in the notes at three stages:

• the note created by the speech recognition software;

• the speech-recognition note after it had been edited by a medical transcriptionist;

• the transcriptionist-revised note after it had been signed by the physician.

The error rate was 7 per 100 words in the notes created by the speech recognition software, falling to below 1 per 100 after transcriptionist review and physician sign-off. In terms of clinically significant errors, 64% of the speech-recognition notes had such errors, which fell to 8% after review and sign-off.

The researchers conclude that their study "highlights the crucial role of manual editing and review" when using speech recognition software.