LAUSR

Automated speech recognition (SR) technology—defined as computer-assisted transcription of spoken language into readable text in real or near-real time—is becoming ubiquitous in everyday life. SR has been already integrated into many electronic devices (e.g., personal computers, mobile phones, smart homes) and is envisioned to revolutionize the way we interact with technology in the near future [1]. In medicine, SR was adopted early in several fields, such as radiology [2], but was not accepted uniformly across all clinical settings. Today however, with the widespread adoption of electronic health records, SR is becoming increasingly prevalent across many types of clinicians in multiple healthcare settings. Previously, SR technologies in healthcare were adopted with caution because of safety concerns and the potential for errors. With the rapid proliferation of SR into different domains of healthcare, only a few studies have examined the safety and accuracy of these systems. For example, a systematic review published in 2016 [3] found that only ten studies to date have focused on the safety of SR systems. Although the accuracy of SR systems has grown over the years [3], their safety is still a significant concern. For example, a recent study has found that an SR system in the emergency room made 1.3 errors per note on average and 15% of the errors were judged clinically significant [4]. Another study conducted in 2017 found that the rate of errors in SR-system-generated clinical notes was more than four times higher than that in non-SR notes [5]. However, our literature review did not find any studies on whether SR errors have led to actual patient harm. To bridge this gap in knowledge, we analyzed a large database of medical malpractice claims to assess patient harm related to SR.