Why Accurate Transcripts Can Still Be Clinically Wrong (Full Transcript)

[00:00:00] Speaker 1: One problem I'm seeing recently that's becoming a pretty interesting problem to solve is typically we've thought about, oh it's LLMs that hallucinate. I think one thing we're seeing though is even though the transcript is correct, let's say assembly did the job, they returned exactly you know what a human might think was said, it doesn't mean it's necessary like it might clinically make sense, which is like so much clinical context that whether it's like this dosage for this medication actually is impossible. So even though it sounded like that's what the patient said or I want one case that happened literally earlier today that I was debugging, it was like the provider said like 20, three second pause then said four. That's two numbers but they actually just meant 24. But the system doesn't know like that sentence may be grammatically correct but it's actually not clinically correct. So what we've started to build is a hallucination layer also on like the transcript and vetting clinical context making sure it's clinically grounded. We've always done that on the note side because that's LLMs will hallucinate way much more or I wouldn't say it's a hallucination but English is not necessarily deterministic all the time. So now we're trying to build safeguards around that. Which is I guess top of mind what challenge we're working through right now.