DeepScribe: The AI-Powered Medical Scribe with CEO Akilesh Bapu

Full episode

Matt Turck [0:54] You brought fans.

Akilesh Bapu [0:57] Yes, I did. Clearly.

Matt Turck [1:29] Wonderful. Well, thanks for doing this. Welcome to this event. So, to say a little bit at the high level, DeepScribe is an AI-powered medical scribe that records the patient visits, transcribes the conversation, and automates the generation of summary notes. So, we are going to talk about all of this, but I'd love to start at the high level and talk about what's happening in healthcare as part of that whole generative AI boom. So traditionally, there's been little bits of AI here and there in healthcare, but it seems to be accelerating, at least from an outsider's perspective.

Matt Turck [1:46] So what's going on? What are you seeing? And I guess, what are the implications?

Akilesh Bapu [2:11] Yeah, absolutely. So healthcare is very much in a time-is-now moment, which is incredible for us as well as other companies in digital health. So, we started DeepScribe—I mean, we came up with the idea in 2018. And back then, healthcare was very different. So, everybody wanted on-prem deployments. No one took AI seriously. And there was a lot of this notion of death by pilot. So, when you went around and tried to sell a product like ours, people would limit it to 20, 25 doctors.

Akilesh Bapu [2:50] You'd be lucky if they paid you for it. And so, at DeepScribe, a lot of our initial journey was actually through private practices. We sidestepped health systems because the sales process and the go-to-market just wasn't there with these larger organizations, because they didn't take us and other healthcare startups seriously. Now, with the release of GPT-4, every human in healthcare is now taking AI super seriously, and everybody knows the practical applications of it, which is a first. And that's something that, as startups, we weren't able to do, but lucky for us, OpenAI was able to.

Akilesh Bapu [3:20] So now there's a stat floating around that 30% of CMIOs, which for us are our customers, are now looking at generative AI solutions and looking to purchase generative AI solutions. So that's a huge first. And for us, a lot of the conversations turned from whether they should adopt ambient documentation to which ambient documentation should they buy. So it's turned from convincing and educating the customers on what generative AI is or what AI is to now convincing them that we're the best product.

Akilesh Bapu [3:55] And for us and for startups, I think that's a lot easier of a sales process to have to deal with. So it's been amazing. The adoption is huge. I was just at HLTH, the premier conference for healthcare and AI and innovation, two weeks ago in Las Vegas. Back in 2019, it was around 500 people in terms of attendance. It was at the MGM Convention Center, which is fairly small. This past year, it was 10,000 people in attendance.

Akilesh Bapu [4:12] So literally 20x in size. But that just shows you how serious healthcare is about adopting innovation now. And the next five years are going to be very, very monumental for what's to come. So it's pretty exciting for us.

Matt Turck [4:42] Yeah, so maybe outside of what you guys do, maybe some thoughts on the opportunities and challenges. Obviously, healthcare, while exciting now, is still a very kind of specific industry with privacy and all sorts of different things obviously very important, sort of like clinical outcomes. So what do you experience and foresee in terms of opportunities and challenges in healthcare in general for AI?

Akilesh Bapu [5:04] Yeah, I think the biggest thing is AI burnout. So back in 2018, when we started the company, we weren't the first to do 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. 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, it was hard for them to take us seriously.

Akilesh Bapu [5:30] I think with AI, it's been a reset 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. Let's take them seriously. Let's adopt them. But it also runs the risk of that same burnout occurring again. So I think it's up to the startups in terms of being super responsible about customer experience and making sure they deliver on the value, because people are definitely buying.

Akilesh Bapu [5:50] But what's not shown yet is whether these solutions can gain true system-wide adoption. So can you expand past 25, 50 clinicians at a health system?

Matt Turck [6:01] 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?

Akilesh Bapu [6:26] Yeah, so before DeepScribe, I was a research scientist at BAIR, which is Berkeley's AI research lab. And our lab was very first principles in terms of AI because it was on the border between the stats department and the engineering department, as well as my mentor, Jamie Murdoch, who had actually written one of the first papers on interpretability when it comes to natural language processing. So we were all about data labeling, data curation, and it sucked for me as an undergrad student because all I wanted to do was train large models on a lot of data.

Akilesh Bapu [6:48] But my mentor took me aside and was like, "We're going to 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.

Akilesh Bapu [7:21] 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 was get him onto the latest software, which at the time was Nuance and their dictation tools. So that was 10 years ago. And for those of you who don't know, documentation for clinicians basically means writing down about one to two pages of what they just talked about with the patient.

Akilesh Bapu [7:56] So this is everything from their stories, their symptoms, their history, all the way to the diagnosis and the thought process behind the diagnosis. And on top of that, they have to put this documentation into electronic health record systems. So these are software that have achieved fairly wide penetration across systems and are not really designed for clinicians. They're designed for payers, which is another major issue. But that paired together means that clinicians spend about three hours a day on this documentation.

Akilesh Bapu [8:20] So it's a fairly large issue. People have been talking about it for at least a decade now. And back when I was shadowing my dad, in a sense, as a kid, it was very much at large. Fast forward about a decade, I would say I was on the cutting edge in terms of AI work at Berkeley. And it was very cool and we achieved a lot, but every time I went home, it was like we were going a decade in the past.

Akilesh Bapu [8:46] So my dad was using the same products he was using 10 years ago, spending the same time or even more time on documentation that he was 10 years ago, and the problem was still very real. So I ended up actually, as a naive ML researcher, trying to build something myself that solved the problem. So I took the same first-principles approach, tried to collect the data, tried to figure out which model would work, and then ship it to my dad.

Akilesh Bapu [9:20] So I could see how much time he was able to save. And that's where I got stuck, is the data. So there was no open dataset of patient-physician conversations on which I could train a solution like this. And that sort of spoke to more of a fundamental issue in healthcare, which is that there was this broken data pipeline. A lot of the source-of-truth information on which healthcare is based isn't really collected effectively and used as a feedback loop back to the patient to influence healthcare outcomes.

Akilesh Bapu [9:47] So that was the kind of bigger problem that we really wanted to solve at DeepScribe. And the first step is to actually collect this foundational dataset of patient conversations. So that's how we started. Because the AI wasn't writing the note out of the box, we actually hired hundreds of contractors around the world to be that human loop. A doctor would record a conversation, we would transcribe it using speech-to-text, we would summarize it to the best of our ability using the stack we had at that time, and then a human would review it, label the data.

Akilesh Bapu [10:18] And we were actually an internal tools company for a long time, so we would basically try to make the human faster at labeling that data, make the quality consistent. At the 5 million mark in terms of conversations, that's when we really started to see the AI make a giant leap in terms of note completion. So today, we actually, a couple months ago, for the first time achieved parity with our AI and our expert human scribes, at which point we took them out of the loop.

Akilesh Bapu [10:46] And so now DeepScribe is fully automated, which is super exciting for us, because in this time-is-now moment where enterprises are now willing to sign system-wide contracts for solutions like ours, we're finally ready to support them. So it's been quite the journey, but it's been a fun one.

Matt Turck [11:13] Yeah. What was that like, by the way? So you've been working on something for years, and then there's this new thing that comes and basically kind of blows out of the water whatever you were doing before. Was it obvious to you? It's like, okay, let's burn all the boats and go and do this. Or was it a little bit of a feeling of, well, we built this whole thing, what are we going to do with it?

Akilesh Bapu [11:38] Yeah, it wasn't obvious at all. So the first gut is always to slow-roll it. So we were like, let's keep it gradual. So as the AI gets better, we'll slowly phase the human out of the loop. But that wasn't the right mindset to have. Because in order to get the human out of the loop, we had to be intentional about it. We had to start building around the fact that a human won't be there anymore. And that meant investing more into the interface for a physician to correct the note and a lot of that post-note-delivery workflow that wasn't there before.

Akilesh Bapu [12:15] So yeah, we felt like a big company immediately, overnight. And a couple of weeks in, I think we were talking as a leadership team, and we were like, what are we doing here? Clinicians are barely sending the note to the human to review, yet we were keeping that option. We were keeping this base-level contract with our offshore suppliers. So we decided to officially make the move. And that's when we started to see a complete shift. So it had to be intentional.

Matt Turck [12:34] And maybe just to drive it home before we get into how it works behind the scenes, maybe give us a quick product tour of what the product exactly does in terms of different modules and different capabilities? And then, yeah, we'll go into how that works.

Akilesh Bapu [12:58] Yeah. So we deeply integrate with most EHRs out there. So before a clinician uses DeepScribe, we sync that patient schedule. So DeepScribe knows all about the patients the clinician's about to see. It's all loaded within their app. We primarily use the iOS app. So clinicians would open up the app, before they go in to see their patient, they would select the patient, provide some context to DeepScribe. So they would summarize their current line of thought and thinking in terms of medical decision-making.

Akilesh Bapu [13:07] DeepScribe would understand that, create a pre-draft.

Matt Turck [13:14] It's literally like you're my doctor, you put the phone on during the visit, and as you and I chat, DeepScribe records.

Akilesh Bapu [13:15] Exactly.

Matt Turck [13:17] Obviously, with my consent.

Akilesh Bapu [13:38] Exactly. So, as a clinician, I would ask for your consent before recording. So I'd be like, Matt, I'd like to use this to help me help you. So if I record this conversation, I'll be able to focus and maintain eye contact with you and deliver the best possible care. And I also won't forget information because my memory is often burdened as a burned-out clinician. So I'd take DeepScribe out, select the patient, put it down, record the conversation, and I talk naturally as I would before.

Akilesh Bapu [14:06] Once the conversation's done, we'll go ahead and process the note. So, I know we'll dive behind the scenes, but we'll transcribe it, we'll summarize it, and then we'll actually show you the note within a few minutes after the encounter so you can go ahead and approve it. Once you do that, it'll go into your EHR and you'll be done with your documentation.

Matt Turck [14:14] So, the various components of this, presumably speech-to-text and then analysis of the text?

Akilesh Bapu [14:15] Exactly.

Matt Turck [14:17] How does that work?

Akilesh Bapu [14:44] Exactly. So the first step is speech-to-text, and that was tricky for us because it was really hard to know which speech vendor at which point had the best performance and what would be the quality of the conversation or the acoustic scenario in which it had that performance. So we actually hired one of the leading speech engineers from Deepgram to build our in-house speech recognition system. That was my decision. I thought that was the way to go.

Akilesh Bapu [15:11] And on day one, he was like, Akilesh, that's the wrong way to go. We are not building in-house speech. That is a complete waste of time. Instead, let me introduce multi-armed bandits. So let's integrate all the speech vendors. We're not going to do any sort of benchmarking or word error rate analysis and pick one. We're going to use several and let the algorithm pick which speech vendor to use for each type of conversation. So we set an optimization function where we picked a metric for it to optimize.

Akilesh Bapu [15:37] So every conversation, it's basically learning how a certain vendor did and defining a weight for that vendor. So at any given point, if Google releases a new version of its speech model, the bandit will automatically learn whether or not it should be using that instead of Microsoft's or Amazon's, or whether it should keep using it.

Matt Turck [15:41] So, and it decides that in real time based on each conversation, or—

Akilesh Bapu [15:43] It decides that in real time. Yeah.

Matt Turck [15:45] Based on what, on the content of the conversation?

Akilesh Bapu [16:02] It decides that based off of the number of edits the clinician makes. So it'll have a certain experimental list of criteria. So if you're like, experiment on 20% of conversations, but pick the best one on 80%, we'll use that 20% to try the newer models that were released. And based off of that, it'll either replace the primary model it uses for that group of conversations, or keep the model it was using before.

Akilesh Bapu [16:16] So it's always learning, and you can define the rate at which it's learning.

Matt Turck [16:20] Okay, great. So that's the speech-to-text part. What's the next bit?

Akilesh Bapu [16:51] So for the actual summarization, we've fiddled with this over time, but the one that gets us the highest accuracy is actually three separate models. Five million conversations right now. They're all labeled. And so basically, we have a stack of classical information extraction techniques paired with our own in-house LLM that we've fine-tuned and are currently in the process of pre-training. And then we have GPT-4 that's also used. And so, depending on the task, we will either use one, two, or all three of them and see whether they agree or not.

Akilesh Bapu [17:09] And by doing that, we have a way to validate the output, but then also leverage the nondeterminism of language models that makes them so good.

Matt Turck [17:34] And you fine-tune them using that? Okay. What have you learned in that effort of fine-tuning GPT-4 in terms of what's—I don't know how easy it is, what works, what doesn't work.

Akilesh Bapu [17:54] So with GPT-4, to be honest, we haven't gotten the impact we'd like in terms of fine-tuning. Where we've seen the most impact is with our own in-house LLM. And with that, I think the primary thing, and this is probably obvious to most people by now, is the quality of the data really matters. Five million conversations—the way we got that to work really well is by continuously monitoring and curating data and labeling it and making sure it's super high quality. It's exactly like the output we want.

Akilesh Bapu [18:31] So that's really helped the most. And obviously, using larger models. So as open source releases more and more parameters, we jump on that. And it makes a big difference, even though some may say it's an emergent phenomenon. That's sort of one of the pillars we like to continue to experiment with.

Matt Turck [18:52] You wrote somewhere about supervised versus unsupervised learning. Can you go into this? Which part of this is still relevant? And maybe to make this interesting for a wide group of people, define supervised versus unsupervised learning.

Akilesh Bapu [19:12] Yeah, I think we have a mix of both. 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 you can kind of call it supervised on some of the steps, like the RLHF step. But for us, we use both in a pretty healthy manner. So a lot of our pre-training is done in an unsupervised manner, but then that supervised phase of learning is still important because we do train it on certain tasks afterwards.

Akilesh Bapu [19:33] So the fine-tuning step, as well as the RLHF step, all the stuff we do with our classical models, is very much supervised. So I think in healthcare for us—

Matt Turck [19:36] Maybe just define what it is for the

Akilesh Bapu [20:01] So supervised is where you typically, for a specific task, give the model an input and an output. So you almost need the gold standard answer for each of those tasks. Unsupervised, you don't exactly need that. So, for example, in the famous Transformer paper, you're training it on typical English language, so you're masking a specific word, asking it to predict that word. So you don't need the end-task labels.

Akilesh Bapu [20:11] So it can be powerful for training on lots of data, but for small amounts of data, supervised is typically best.

Matt Turck [20:15] Great. And you mentioned automation.

Akilesh Bapu [20:15] Are you—

Matt Turck [20:19] Do you still have humans in the loop somewhere or not?

Akilesh Bapu [20:44] So for healthcare, I think one of the biggest drivers of adoption is trust. And so for some of our enterprise customers, we still have that human as a final step. But what we've done is we've essentially let the doctor choose whether or not they want that human. So for any given note, they can send it for human review. It takes a little bit longer to get back, but some of them prefer it rather than editing it themselves. But for most of our customers, they're now on the fully automated solution.

Matt Turck [21:01] Okay. You announced recently, at the beginning of this month, the Customization Studio. What is that?

Akilesh Bapu [21:32] Yes, so Customization Studio is probably the most important facet of DeepScribe's product. So in this AI age, I think it's fairly easy now to record a conversation and generate a note with GPT-4. But what really makes DeepScribe different from a lot of those solutions is the ability to conform to nuanced workflows for clinicians, 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 transform their note to how they like it.

Akilesh Bapu [22:06] So we can natively fit into most of their workflows. We can collect discrete fields from their conversations. We can 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 this interface now allows them to, with a few clicks of a button, change how they like the note. And a lot of it is enabled by some of the advances we've seen in LLMs that allow you full control over the style of language, which has been the big breakthrough that enabled Customization Studio.

Matt Turck [22:30] This is really super cool. So maybe dig into it, like just reviewing my notes. There's something called Progressive Notes. Yeah, there's Blueprints, physical exam. Like, just maybe go into what those different parts do.

Akilesh Bapu [22:56] Yeah, absolutely. So if you look at medicine, the primary reason we chose primary care as the first customer is because all of that information comes from the conversation. So we record the conversation, write a summarized note. That's fairly boilerplate, and it works out of the box. But if you look at oncologists, about 50% of their information comes from the conversation, about one-fourth comes from a summary that they would have done, and then another fourth may come from previous visits. So we need a way to incorporate that information when it comes from all these different sources.

Akilesh Bapu [23:29] So Progressive Notes, like you mentioned, allows you to basically pull information from previous visits. The AI is basically going through all the different charts a clinician would have written in the past on that patient, learning and creating a knowledge state, and then using that to influence how it summarizes the current note. Blueprints is especially helpful for primary care physicians that are doing annual wellness visits, and I think pretty much everybody in this room probably would have done that with their clinician. And in that, there's specific discrete information that's important—so when the last visit was, what their patient's height is, what their blood pressure was, did they take a COVID vaccine or not, if so, which one.

Akilesh Bapu [24:03] So we can actually define those things with AI so it can pull it from the conversation. And if you look at DeepScribe's product, it has a summary and then it has the discrete information at the bottom. So this allows you to pull in that discrete information. But for the clinician, what this really means is now we're able to support most specialties out there regardless of how nuanced the workflow is. So back to this enterprise wave of customers looking to buy solutions like this for all their providers, the specialties you support really matters in terms of how much you can penetrate that organization.

Matt Turck [24:22] The other major question that's always on buyers' minds in the enterprise is the issue of hallucinations.

Akilesh Bapu [24:23] Yeah.

Matt Turck [24:30] And you guys have done some very interesting work around this. Do you want to talk about your evaluator and fixer LLMs?

Akilesh Bapu [24:47] Yeah. So one of the things I love about language models is the flexibility with which you can ask them to do certain tasks. So we actually ask the language model itself to look at the output and then fix itself if it notices any issues with it. And over time, we've gotten fairly clever with the prompts, fairly clever with how we fine-tune the model so that it's able to not just predict the task but also assess the output of itself.

Akilesh Bapu [25:05] So that's really crucial for avoiding hallucinations. But the other thing that really helps is having—

Matt Turck [25:10] Just to drive this home, it's like AI fixing AI, right? It's like AI looking at the output.

Akilesh Bapu [25:36] Exactly. And then the other thing that really helps with hallucinations is validation models. So for a given task, we'll have a classical model typically produce that same output. And then if the LLM puts in something that the classical model didn't have in its output, we'll go ahead and only go with the classical model's output and disregard the LLM's output. So that's really helped with validation as well.

Matt Turck [25:44] So much cool stuff. You also have—there was a preview of DeepScribe for Vision Pro.

Akilesh Bapu [25:47] That was actually a joke.

Matt Turck [25:52] That was a joke. Okay. That would be so cool. That's definitely not part of the roadmap.

Akilesh Bapu [26:09] That would be super cool. It depends on Apple's Vision Pro adoption across its customer base of health systems, which probably will never happen, sadly. But no, our designer sent us that as soon as Vision Pro came out. Okay, very cool.

Matt Turck [26:31] What are you learning on the sort of go-to-market side of things? So you mentioned you had started with private practices at the beginning of the company. But the ecosystem, the industry of healthcare has evolved now to embracing AI. So where are you in that journey? Are you still selling to—who are you selling to and how? And how are they reacting?

Akilesh Bapu [26:58] Yeah. So in the beginning, it was really funny. So my co-founder and I, this was our first job out of college. And so we were naive students, low 20s in terms of age. And the first sales motion for us was getting on his Vespa and riding around San Francisco, pitching DeepScribe to every single hospital we could find on Google Maps and seeing what they thought and if they would buy the product. We actually ended up in the red that day because my co-founder got a parking ticket.

Akilesh Bapu [27:32] So that didn't work. And so we started to take a step back and figure out how we can hack the sales process here, how we can hack distribution, because digital health sales has just historically been one of the hardest places to sell into. And then we took a step back and asked ourselves, like, how do we buy products? We look at Facebook, we search for it, we typically make that search as soon as we experience the pain point. And so with doctors, because we would actually bucket them more in the consumer space rather than the B2B space, specifically for private practices.

Akilesh Bapu [28:03] We asked the same question to them, and they were like, well, I'm scrolling through Facebook at night when I'm tired of my documentation. I use Instagram. I've even searched for scribes. And we were like, okay, why isn't anybody using inbound-based marketing for digital health solutions? And so we actually became one of the first to do that. We bought AdWords on Facebook and Google, and all of our traffic and all of our initial sales from, I think, $0 to $6 million in ARR came from inbound.

Akilesh Bapu [28:39] So no cold calling, no cold emailing. It was all through inbound. And that worked really well with private practices. And that was our go-to-market hack that worked extremely well. With enterprises, obviously it's a little bit different. We were riding the inbound wave for a little bit because the CMIO or a clinician from a health system would also be searching for those terms, would find out about DeepScribe that way. But it wasn't a very deterministic way of selling to enterprise.

Akilesh Bapu [29:10] So we actually bit the bullet and now have our own outbound sales team. So enterprise sales has been fairly classic in terms of having your fair share of SDRs and AEs and ABM. So no hacks there, unfortunately. The only hack is that we really ride off the backs of Nuance, which is the 800-pound gorilla in our space that is trying to do something similar to what we're doing, but they have the distribution advantage. So they've gone and educated all the customers, and we follow on and basically say our product is better.

Akilesh Bapu [29:21] And that's shown to be fairly effective to date. Okay, very nice.

Matt Turck [29:51] Obviously, in healthcare, there is a whole iceberg around regulation, privacy, all the things. How do you all think about it? You're taking patient data, you're moving it to the cloud somewhere, sending it possibly to OpenAI. How do you safeguard the whole process?

Akilesh Bapu [30:18] 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 when a patient basically fills out their initial paperwork, they'll see a diagram of our data pipeline and what we're doing with it, what models we're training with it. 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.

Akilesh Bapu [30:46] 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 100% consent from patients, which was a big step because initially, especially because a lot of the early users of DeepScribe were in rural areas and rural communities, patient consent was pretty difficult for us, but that helped dodge that. In terms of privacy in general, I think right now we are good as long as we don't try to replace the clinicians, which we don't expect to do or don't intend to do.

Akilesh Bapu [31:15] We want to be that trusty assistant that ends up getting the clinician's approval at the end of the day. So that's how we think about privacy and regulations right now. So while it's a consideration, it doesn't affect the go-to-market right now, at least. Great.

Matt Turck [31:38] So maybe one last question from me, and then I'll open it up to you all. Just zooming out, what other areas of healthcare do you think are ripe for AI disruption or enhancement? If you were not doing this, what else would you be looking for?

Akilesh Bapu [31:58] Yeah, so I'll start with DeepScribe's vision first so I can show what's in scope and what's not in scope. So our vision at DeepScribe, the reason we started with scribing is because we get the foundational patient dataset to solve scribing, but then also solve some of the bigger problems in healthcare. We want to be the platform for truly end-to-end AI-based care in the future, which means that DeepScribe will eventually become an extension of the clinician, with the end goal of being a trusted copilot that not only records the conversation and writes the documentation, but gives that bidirectional feedback back to the clinician to end up improving care.

Akilesh Bapu [32:38] And this is aimed at the large issue in medicine where about 5% of patients on average are misdiagnosed, and medical errors are known as the third leading cause of death. So this data feedback loop is very important, and we want to complete it. So that's DeepScribe. But in terms of—

Matt Turck [32:42] There's nothing left. You've just conquered all of healthcare. Exactly.

Akilesh Bapu [33:14] That's why I started with that. But in terms of areas that I would encourage folks to tackle, I think one of the things that has really, really been on my mind recently, as I go back and forth with my dad, who's an oncologist, is the availability of cheaper, more accessible testing. The way insurance coverage of testing works is fairly screwed up, in the sense that it's all probability-based. And so, if a patient has a 98% chance of not having a condition, then they just won't advise testing.

Akilesh Bapu [33:47] But that means 2% of those patients end up with that condition, and that usually ends up being fatal. So that's something that I think deserves more exploration. And I think with AI, with the cameras on cell phones that are available today, it could be very, very interesting what you can do with testing.

Matt Turck [33:50] All right, very cool. Thank you so much. Really enjoyed it.

Akilesh Bapu [34:23] Really appreciate it. Thanks. Thanks for joining us for The MAD Podcast. We're back here every Wednesday with new conversations with leaders in the machine learning, AI, and data space. And if you like this show, you can also find a video recording of not only this episode, but many, many more over on the Data Driven NYC YouTube channel. Thanks again, and catch you next week.