The Exchanges

Q What was your journey into starting the company? I read somewhere that your dad is an oncologist, like you grew up in a health family. What was your journey?

A Yeah. Uh, so before DeepScribe, um, I was a research scientist at Bayer, which is Berkeley's AI research lab. Um, and our lab was very first principles, um, in terms of AI, because it was on the, um, on the border between the stats department and the engineering department, as well as my mentor, Jamie Murdock, who had actually written, uh, one of the first papers on interpretability when it comes to natural language processing. So, um, we were all about data labeling, data curation, Um, and it sucked for me as an undergrad student, because all I wanted to do was train large models on a lot of data. Uh, but, you know, my mentor took me aside and was like, we're gonna start small, and so that was my background. The way I actually got pulled into healthcare was through my dad, who's an oncologist, and I would say as a kid, I was actually desensitized by the problem of documentation, because I just assumed it was a way of life for my dad to spend the evenings catching up on notes, or on the weekends, missing important Life events because he had to, he was hitting that seven day mark in which health systems required you to complete your documentation by. So as a kid, couldn't really do anything about it. Just accepted it. All I could do is, um, get him onto the latest, uh, software, which at the time was Nuance and their dictation tools. Uh, so that was 10 years ago. Uh, and for tho…

AI assessment note: “The way I actually got pulled into healthcare was through my dad, who's an oncologist”

Q Okay, great. So that's the speech to text part. What's the next bit?

A Um, so for the actual summarization, um, we've, we've filled with this over time, but the, um, the one that gets us the highest accuracy is actually three separate, um, three separate models. So the first is our own classical models that have been trained on all of our data. And we have a little over two and a half million conversations right now. They're all labeled. And, um, so basically we have a stack of classical information extraction techniques, um, paired with our own in-house LLM that, um, we've fine-tuned and are currently in the process of pre-training. And then we have our, uh, we have GPT-IV that's also used. And so, um, depending on the, the task, we will either use one, two, or all three of them and see whether they agree or not. And by doing that, we have A way to validate the output, but then also, um, but then also leverage the non-deterministicness of language models that makes them so good.

AI assessment note: “for the actual summarization, um, we've, we've filled with this over time”

Q Um, you announced recently the beginning of this month, uh, the customization studio. Uh, what is that?

A Yes. Um, so customization studio is, um, probably the most important facet of DeepScribe's product. Uh, so in this AI age, I think it's fairly easy now to record a conversation And generate a node with GPT-IV. Um, but what really makes DeepScribe different from a lot of those solutions is the ability to conform to nuanced workflows for clinicians. Um, especially when it comes to the higher revenue generating folks like specialists, high patient volume, because for them, every single second of documentation time matters a lot. So customization studio gives clinicians about 35 different ways to, um, transform their note to how they like it. So we can natively fit into most of their workflows. We can collect discrete fields from their conversations. We can, um, change the style of the writing, and we put that all into clinicians hands. So previously, In healthcare, clinicians haven't really had a good way to configure and train their own models without it looking like a black box. So, uh, this interface now allows them to, with a few clicks of a button, uh, change how they like the note. And a lot of it is enabled by, uh, some of the advances we've seen in LLMs that allow you full control over the style of language, um, which has been, uh, the, the big breakthrough that enabled customization studio.

AI assessment note: “customization studio gives clinicians about 35 different ways to, um, transform their note”

Q guys do, um, Maybe some thoughts on like the opportunities and the challenges, obviously healthcare while excited now is still a very kind of specific industry with, you know, privacy and all sorts of different things, obviously very important, uh, sort of like clinical outcomes. Um, so w w w What, what do you for experience and foresee in terms of opportunities and challenges in healthcare in general for AI?

A Yeah, um, I think the biggest thing is AI burnout. So, um, back in 2018, when we started the company, we weren't the first to do, um, any sort of ambient AI documentation. There were a couple others, and because the state of AI was very different, the quality of products really mattered back then. Um, clinicians and the customer got burned out by solutions that didn't exactly solve the problem. So when we came and said that we solved that problem, except better, um, it was hard for them to take it seriously. I think with AI, it's been a reset, um, for them. So now they have shifted back and they're like, tech is actually a lot better now. So these solutions can actually solve the problem for the first time. Um, let's take them seriously. Let's adopt them. But, um, it also runs the risk of that same burnout, uh, occurring again. So I think it's up to the startups, um, in terms of Being super responsible about customer experience and making sure they deliver on the value, because people are definitely buying, um, but what's not shown yet is whether these solutions can gain true system-wide adoption. So, um, can you expand past 25, 50 clinicians at a health system?

AI assessment note: “I think the biggest thing is AI burnout.”

Q that comes and that basically kind of like blows out of the water, like whatever you were doing before. Like, what did you, um, was it obvious to you? It's like, okay, let's burn all the boats and go and do this. Uh, or, uh, was it like a little bit of a feeling of, uh, well, we build this whole thing. What are we going to do with it?

A Yeah, um, it wasn't obvious at all. So the first gut is always to slow roll it. So we were like, you know, let's, let's keep it gradual. So as the AI gets better, um, we'll slowly phase the human on the loop. Um, but that wasn't the right mindset to have. Um, because in order to get the human on loop, we had to be intentional about it. We had to start building around the fact that a human won't be there anymore. Uh, and that meant investing more into the interface for a physician to, like, correct the note. Um, and a lot of that post note delivery workflow that wasn't there before. Um, so yeah, we felt like a big company immediately, like overnight. Uh, and a couple of weeks in, I think, uh, we were talking as a leadership team and, um, we were like, what are we doing here? Like, we, You know, clinicians are barely sending the note to, to the human to review, yet we were keeping that option. We're keeping this like base level contract with our offshore suppliers. So we decided to, um, officially, uh, make the move and that's when we started to see complete shift. So it had to be intentional.

AI assessment note: “it wasn't obvious at all. So the first gut is always to slow roll it.”

Q Uh, what, what have you learned in that effort of fine-tuning GPT-IV in terms of like, what's, I don't know how easy it is, what works, what doesn't work?

A Um, so with GPT-IV, to be honest, we haven't gotten the, um, the impact we'd like in terms of fine-tuning. Uh, where we've seen the most impact is with our own in-house LLM. Um, and with that, we, I think the primary thing, and this is probably obvious to most people by now, is the quality of the data really matters. Um, so even if we trained it on two and a half million conversations, the, the way we, Um, got that to work really well is by, uh, continuously monitoring and curating data, um, and labeling it and making sure it's super high quality. It's exactly like the output we want. Uh, so that's, that's really helped, um, the most, and obviously, um, uh, using larger models. So as open source releases more and more parameters, we jump on that, um, and it, it makes, uh, it makes a big difference, um, even though, Uh, some may say it's emergent phenomenon. We, uh, that's, that's sort of one of the pillars we like to continue to experiment with.

AI assessment note: “with GPT-IV, to be honest, we haven't gotten the, um, the impact we'd like”

Q You wrote somewhere about, uh, supervised versus unsupervised learning. Um, can you go into this? Where, where, where, where, you know, which part of this is still relevant and maybe to make this interesting for, uh, you know, what group of people define supervised versus unsupervised learning?

A Yeah, um, I think, you know, we have a mix of both. Um, I think the general, I know the community changes their stance on this every now and then, but I believe people still consider LLMs as unsupervised still, although there's, uh, you can kind of call it supervised on some of the steps, like, um, uh, like the RLHF step, but, uh, for us, we use both in a pretty healthy manner, so a lot of our pre-training is done in an unsupervised manner, uh, but then that supervised, uh, phase of learning is still important, Um, because we do train it on certain tasks, um, afterwards. So the fine tuning step, as well as the ROHF step, all the stuff we do with our classical models, um, is very much supervised. So I think in healthcare for us...

AI assessment note: “a lot of our pre-training is done in an unsupervised manner”

Q Very nice. Um, Obviously in the healthcare, there is a whole, you know, iceberg around regulation, privacy, uh, all the things. Um, how do you all think about it? Um, you know, you, you taking, uh, patient data, you moving it to the cloud somewhere, um, sending it possibly to open AI. Um, how do you, how do you, how do you safeguard the whole, um, process?

A Yeah. So patient consent for us is number one. So we want the patients to know every single thing we're doing with the data. So, uh, when a patient basically fills out their initial paperwork, they'll see a diagram of our data, uh, pipeline and what we're doing with it, what models we're training with it. Um, and even if it goes over their head, it's nice to share with them that information and the clinician basically double clicks on that. So they go to the patient And they're like, help me help you. Um, and that really helps the patient also fully understand what they just signed up for. And through that, we've been able to get basically a hundred percent consent from, from patients, um, which was a big step because initially, um, especially because a lot of the early users of DeepScribe were, um, rural areas and rural communities, um, patient consent was pretty difficult for us, but that, that helped. Dodge that. Um, um, the, in terms of privacy in general, I think right now, um, we are good as long as we don't, um, try to replace the clinicians, which we don't expect to do or don't intend to do. Um, we want to be that trusty assistant. Um, that ends up getting the clinician's approval at the end of the day. So that's, that's how we think about privacy and, and regulations right now. So that's, um, while it's some, while it's a consideration, it doesn't affect, uh, the go-to…

AI assessment note: “patient consent for us is number one. So we want the patients to know”