The Lack of Explainability in Automatic Speech Recognition Can Cause Faux Data Work

Abstract

Automatic Speech Recognition (ASR) is increasingly used in healthcare to reduce documentation workloads by transcribing spoken words into Electronic Health Records (EHRs). However, these systems, based on machine learning, require ongoing data annotation and validation by healthcare professionals to ensure accuracy. This paper, based on fieldwork at a public Danish hospital, investigates the challenges healthcare professionals face in detecting and addressing technical issues, such as glitches, within ASR systems. Using mixed methods, the study reveals that healthcare professionals spend significant time annotating and training the machine learning algorithms—time that could otherwise be dedicated to patient care. Without access to clear metrics, like recognition rates, healthcare professionals are unable to effectively evaluate their data annotating efforts, leading to "faux data work," where data tasks seem productive but fail to improve system performance. The paper proposes two strategies to mitigate this issue; 1) providing transparent system metrics to enhance user engagement; and 2) creating structured sites of collaboration between healthcare professionals and IT professionals for better reporting of technical issues. These solutions aim to reduce inefficiencies and improve ASR accuracy in clinical settings.

Keywords: Automatic Speech Recognition, Faux Data Work, Transparency, Glitches

DOI: 10.54941/ahfe1005835