By Adele Field, ZyDoc Marketing & Communications Director
Anyone who has tried to seriously use speech recognition in the last few years for clinical documentation – or any other purpose – knows two things: SR is far from perfect, and SR is getting better all the time.
It’s popular in emergency departments for obvious reasons. The concept is a natural match for the rapid pace of an ER or radiology department – it allows doctors to transmit critical information quickly so that more patients can be seen, diagnosed more quickly, and treated more effectively.
Speaking is much faster than keyboarding. Using front-end SR (dictation with no editorial intervention by a third party on the back end), the near-instant conversion of speech to text is faster than dictating for later editing or transcription for entry into the EHR. But speed over accuracy is never a comfortable trade-off in healthcare. In non-emergency clinical settings poor quality documentation is even less acceptable. And the speed of front-end SR is deceptive. Compared to a short delay for transcription or back-end editing, the front-end method loses some of its edge when the time and frustration required for error corrections are factored in. An experienced human transcriptionist or editor will produce much better results In terms of accuracy, with the bonus of expedited STAT turnaround as needed, and automated section-level EHR entry.
Time considerations aside, accuracy is a big deal, whether in the ER, a hospital, or clinic. A mis-captured term or drug name, not discovered, can seriously impact patient safety. We’ve all laughed or been horrified at bizarre text output from an SR app. As an app learns a speaker’s voice and phrasing, these errors are theoretically reduced, producing increasingly more reliable results. Even so, trusting SR without human supervision could be likened to leaping off a high cliff at night with a parachute that intermittently works. Or doesn’t.
On balance, is speech recognition safe? One experimental study, published in JAMIA in 2017, of SR use by 35 emergency department clinicians showed "significant increases in the occurrence of all classes of PPH [Potential for Patient Harm] errors when using SR across both task types: major PPH simple task ..., complex task ..., moderate PPH simple task ..., minor PPH simple task ..., and complex task.1
That the stakes are huge is not subject to debate. At the top of the list are patient health, privacy, happiness, quality of living, life itself, and death. Minor errors that don’t affect patient safety may be relatively tolerable; critical errors could be disastrous.2 Quality Assurance for clinical documentation is at least as important as speed, even in an ER or radiology setting.
So the answer to whether SR is safe enough is "maybe." The only way to reduce risks associated with front-end speech recognition errors is to mandate immediate careful and detailed review. On the other hand, transcription and edited SR are more reliable from the outset – the industry minimum standard is 98% accuracy – and they are comparatively fast when all factors are considered.
Efficiency and safety of speech recognition for documentation in the electronic health record.
Tobias Hodgson, Farah Magrabi, Enrico Coiera
Journal of the American Medical Informatics Association, Volume 24, Issue 6, 1 November 2017, Pages 1127–1133, https://doi.org/10.1093/jamia/ocx073
Published: 27 July 2017
2 AHDI definition:
A critical error is any error in a patient care record that has the potential to:
1. Adversely impact patient safety.
2. Alter the patient’s care or treatment.
3. Adversely impact the accuracy of coding and billing.
4. Result in a HIPAA violation.
5. Adversely affect medicolegal outcomes.
Source: Association for Healthcare Documentation Integrity (AHDI), www.ahdionline.org.
“Healthcare Documentation Quality Assessment and Management Best Practices” https://c.ymcdn.com/sites/www.ahdionline.org/resource/resmgr/toolkits/QABP17_Error_Values.pdf