Webinar Wednesday – AI Ambient Scribes – the Lower-Risk Use Case of Generative AI
Generative AI is taking the world, and healthcare, by storm. The accurately predicted textual outputs based on prompts can induce “wow” moments. But there needs to be an industry-wide consideration for how to iteratively allow innovation powered by large language models to be rolled out within the sensitive domain of patient care while minimizing risks. In this presentation, I will review the core technology power of Generative AI and LLM and describe their lower-risk use case of AI Ambient Scribes, compared to other use cases. I will describe how AI ambient scribes are currently tasked with only generating medical notes based on the verbalized content of a clinician-patient audio-recorded dialogue. Since they are not tasked with predicting diagnoses or therapeutics, clinician-users of AI ambient scribes are trained to expect dictation-like outputs from it, and not clinical decision-support recommendations. As erroneous generated medical note content is either due to sub-optimized speech recognition or language processing performance, the approach to clinician review procedures of the AI-generated notes is critical to minimizing overall risk to patients; key features of this approach will be highlighted. Thereafter, two hypothetical examples of the current risk in using ambient scribes will be presented to illustrate certain edge-case challenges with these tools, namely:
- Confabulations/hallucinations by generative AI, i.e. unexpected outputs that are completely unlinked to the prompt content, can in very rare circumstances give rise to an erroneous documented diagnosis or treatment plan, which may appear as a “decision support-like” diagnosis or treatment recommendation.
- In time, if AI ambient scribe accuracy becomes extremely performant, say with 99% of patient encounters, this may lead to clinicians becoming habituated with only doing quick reviews of the outputs, and thus it is possible patient harm may result from imprudent, error-prone clinician review of the 1% of patient cases with erroneous AI-generated content.
- Finally, considerations for how to proactively prevent harm from these two examples will be reviewed. Interaction through simulated cases using Dr. Crampton’s proprietary AI ambient scribe product – AutoScribe by Mutuo Health – will be a key feature of this presentation, as well as a robust amount of time left over for questions of the presenter.
