Shiv Rao, CEO demos Abridge

Let me start with a really quick background on myself. I'm the founder and CEO, but prior to this I worked at a really large health system on the innovation side of their portfolio at UPMC. I was in charge of the provider-facing side of what we were investing in, into startups and large companies for R&D, but also what we were overseeing from an R&D perspective at Carnegie Mellon University with their AI department. A lifetime ago I went to Carnegie Mellon, but in the middle I became a cardiologist. I still see patients. At this point in time, once a month I'll sign up for the shift nobody else wants. It's a privilege for me because I get to stay close to the mission of the company and I also get to dog food Abridge myself. I get to use the technology we're building, and hopefully that's an advantage for us where we can build, measure, learn that much faster with real user feedback.

We were fairly... I think we had a very strong idea of what we were working backwards from, and that's reflected even in terms of seed pitch decks that we put in front of VCs for that first round we raised in early 2019. We had this idea that healthcare is really all about conversations between professionals on one side of the room and patients on the other. It's happening via all sorts of modalities. We couldn't have predicted what would happen with telemedicine during the pandemic, but certainly we anticipated omnichannel being really important, the ability to serve conversations wherever they're happening between those two types of people. We had this thesis that healthcare is about people and outcomes on the right time horizon. What we wanted to do was position our offering as something that could serve all the different constituents in healthcare, starting with the end users. We started with the consumer side because the barrier to entry and the minimum viable threshold for something that's viable for a clinician like me to use is rather high. The technology that you probably used in 2019 or 2018 when you checked us out is still out there. It has over 450,000 people using it. You can download it off the app store, and it's kind of like transcription on steroids, with a little bit of medical steroid, in that you can record a conversation as a patient with permission and then you'll get a transcript of just the steps.

But as we were serving patients with that solution, we were all the while doing the R&D on the provider side. That moment you're speaking to, that's all the rage right now. Over the last year and a half, we've been translating a lot of that R&D into a solution and putting it in front of the enterprise. Now we're connecting the dots where not only can the enterprise serve their professionals, their doctors and nurses, with something that can unburden them from their clerical work, but we can also layer in that patient offering. The same patient offering you used, but it is different when it's flowing from the clinical side. I'd also point out that in the world of AI and speech recognition and large language models, four years is a hell of a long time.

Let me set it up and I'll be a patient. Absolutely. I think we can lead off with just a little bit of context. Healthcare conversations are different from monologue. Dialogue is very different. When I started seeing patients as a doctor over a decade ago, I was dictating a lot of the encounters I had with people or a lot of the procedures I did. I picked up a Dictaphone and you learn how to dictate really efficiently. I would say, '30-year-old male with a past medical history of diabetes and hypertension presents with chest pain. Chest pain began two weeks ago, worsened by this, alleviated by this, they've tried this and that, and now they also have associated symptoms of fevers, chills, nausea, and vomiting.' But I'm saying, 'Next line, exciting past medical history, capital P, past medical...' This is conversation, this is dialogue between two parties. In the dictation world, I always chose English as my language. In the dialogue world, you have to meet the patient where they are. That's where we do a lot of work on multilingual models. There are people at the University of Kansas Health System, for example, who are using Abridge with all Spanish conversations and creating an English note on the other side. There are any number of other languages beyond English that we can also handle. It's multilingual, but it's also omnichannel. If we were integrated with Zoom right now, I'd just hit a button and then you'd start to create the English note. But let me show you the way I used it a few weeks ago.

I have my back turned towards you. Every patient says yes to that idea. Cool. Then I just hit the record button. If you don't mind, maybe we can just have a conversation. Okay, cool. So hey Matt, you're my patient. I heard you're having some chest pain. What's going on?

Okay, when did this all start?

Okay, tell me a little bit more. Did it start when you were doing something active?

Okay, how long did it take till you felt normal again?

I see. Any radiation of that pain to your arm or to your jaw?

Okay, all right. So you know, I'm looking at my medical record right now and you have a history of diabetes and hypertension. I bring that up because this is chest pain and obviously I'm a cardiologist. The first thing I think about is, could this be a heart attack? Could this be what we call unstable angina, which means that you could have blockages in your heart that we need to address immediately? You have these risk factors for heart disease like diabetes, hypertension. Your great-great-grandfather had a history of heart disease too. So I think we should take this seriously. Come in this afternoon, let's get an EKG and...

I'm a cardiologist, so I want to talk about your other problems really quick as well. Your diabetes is out of control. Your last hemoglobin A1c was through the roof, it was like a billion in January. So I'm curious, are you taking your metformin every day?

Yeah, I don't know about that. I think we're going to have to get a nutritionist involved because it's not just the cake. I saw you at McDonald's last week, you had like 20 Big Macs in front of you and a whole bunch of french fries. So I'm going to ask our nutritionist to call you next week, start you on a low-salt, vegan, diabetic diet. It's going to be delicious. But also, I really need you to take Metformin 500 milligrams twice a day. This is so important. And finally, it's time for us to talk about insulin. So our nurse is going to call you on...

We'll get a social worker involved and they'll call you next week and we can make sure that you've got a supply of everything. Okay?

All right, awesome. So we just had a conversation and now I can talk through what happens. The tech goes through every word that I said and you said. It first makes a prediction about what kind of word this was. It needs to recognize the words. If we talked in a different language, and if we interspersed English and Spanish, or English with Haitian Creole, or Brazilian Portuguese, or Hindi, or whatever, the tech needs to be smart enough to recognize all these different languages.

They need this to be documented. Finally, would the patient want this part of the conversation where maybe I used a metaphor to describe heart failure? Maybe I talked about how the heart's like a house and I went off for 20 minutes trying to teach you. Regardless of who the constituent is, what we do is abridge the conversation. We're pulling out the bits that any of those three constituents would want to see documented. If we talked for 20 minutes about the Barbie movie or about politics, something controversial, you can rest assured as a clinician to build rapport however you see fit, knowing that the tech is going to pull out what needs to be in the record. Once it's pulled it out, it then classifies it. When I think about classification challenges, my mind immediately goes to computer vision companies that are classifying a malignant nodule versus a benign nodule.

We need to classify information against, but then finally restructure that data where it makes sense to structure it so that we can slot it into the right workflows. We'll summarize it, we'll write a note. Let me share my screen so that we can go through the note, which was done in seconds. Here's our conversation. For anyone who's not familiar, there's an archetypal way that all doctors, nurses, medical students write notes, not just in the United States but really around the world. There's a whole history behind that that we can get into as well. It's pulling out the information from our conversation and writing those sections out for us in prose where it makes sense. But as I'll show you if there's time, we also structure data.

Yeah, exactly. Let's look at this. The patient presents with recent onset chest pain, tightening, shortness of breath, began yesterday while walking up a flight of stairs, lasted for 20 to 30 minutes. Some light activity also brought it on. You also have some associated palpitations and lightheadedness, prompted them to sit down, didn't pass out, and some pain radiating to the left arm. Has a history of diabetes and hypertension. Your A1c was high, so it didn't believe my joke of a billion. You've been on Metformin 500 twice a day.

This weekend, citing concerns about the cost of the medication. Now, in the assessment and plan, this is problem-based. This is the way primary care clinicians, internal medicine doctors, and even surgeons ideally document. What we're doing is training our models to work backwards from gold standard Medicare. What would Medicare want to see? We're summarizing all the information that Medicare would want to see, that an auditor would want to see, and certainly it includes the next steps: EKG, echo, CBC, BMP, troponin level. In the diabetes piece, A1c was high, nutritionist starts you on a low-salt, vegan, diabetic diet, take your metformin, and we're going to talk about insulin. But let's say we had that conversation and I as a clinician don't remember that piece where you cited concerns about cost.

In minutes again, the evidence shows up. Not only do I see it, I could even listen to it. 'How long did it take? Felt normal again.' That's relevant because we're pulling out, or we're retro-engineering, transparency into large models that don't inherently have them. If you've ever played with ChatGPT, for example, you might not know the whatever percent of the time where it's hallucinating, where it's making something up that wasn't in the data. It's so important to build trust through transparency. Here, for example, we're pulling out a bunch of conditions and procedures, and you can similarly map. When we demoed before, we also showed you a whole bunch of stuff in relation to those back office workflows. We can similarly give you that evidence.

I hit stop, I swivel my chair, and my note is there inside the medical record. It's sitting there in all the right fields. It's not just one sort of unstructured note; it's not like a copy and paste. It's all in its right place. A deeper integration with the medical record also affords the ability to not just have the note in the right place but to exploit what we're also doing: structuring data, pulling out diagnoses, mapping to ontologies that then map to the medical record entities and attributes for all the next steps. If I prescribed Metoprolol 25 milligrams twice a day, Metoprolol would be an entity, 25 milligrams would be an attribute, and twice a day would be another attribute. That's what we can...

As good as that demo looked, we always want the clinician to trust and verify, look at the note as if their intern created it for them. They're very quickly trusting and verifying. Maybe they need to do a quick literature search and see what the latest evidence is on something before they finalize or crystallize their opinion. We want to make sure that clinician is in the loop, whether it's on the note, on orders, on the problems, or any other aspect of the workflow.

Specialties inform the style of the note. Certain sections of the note are a little bit more art and stylistic. You go through your training, your residency, and you learn that as an oncologist, the history lives in the HPI and this is the format it adheres to. It requires a deep integration with the EMR to be able to summarize that. Similarly, if you're a cardiologist, you're probably writing that history around the chest pain in the way we just demoed: talking about the pain, how long did it last, does it radiate, what makes it better or worse. There's a certain style to breaking down any chief complaint a patient might present with. So there's art here, and it's the art aspects where people might still need to edit based on who they are.

It's one thing to be able to create a draft of a note, and there's a whole spectrum of quality on that. There are some notes that you could probably just put in front of a person and they're willing to make a lot of edits. There are other notes that are going to be closer to what you need to see or what gold standard would be in the medical community. Then there are other notes where there's another tier of note that we're really excited about, where we're also building trust into these notes with transparency. You can see where the evidence came from, and it's that much better of an experience. Interestingly enough, a quick side story: there was a clinician at Brown University recently who was using the technology and sent us a message afterwards saying...

A quick check on the back end, obviously in a privacy and security-informed way, we saw that this was discussed and the clinician just hadn't heard that part of the conversation because we're all fallible. We only capture so much of the conversation as it happens. Building that trust where you as the clinician can highlight that part, the hip replacement, and see where it came from, and even potentially listen to your patient's voice when they shared that with you, we think goes a long way in helping build a product that's really enterprise grade. But the other aspect of enterprise grade is what you're talking to: those deep workflows. That requires a level of API access that to date hasn't really been out there. So let's talk about that deep workflow.

It's a game changer for us because it makes all of our dreams come true in terms of our ability to deliver the best possible experience to our clients. A very significant first principle of this relationship is that it's almost as if we are in a lab, trying to figure out how to innovate, where can we take this technology together, and how can we make those cycles faster. Whether it's in a workflow that we might not have previously imagined this technology being able to sing, and finding out how hard it would be to build that connection and then test it at one of our client sites, or if it's more in relation to an implementation that we...

Here, the care delivery experience for people who are using this right now, and it's not presentation, this is in practice. Clinicians tell us that it just makes everything feel different. They feel more present, they feel more alert, they're listening more as opposed to multitasking at the same time. However long of a game this is going to be to demonstrate that outcome improvement, we think that will be the most profound thing we can demonstrate with technology: that it improved the care delivery experience for the patient.

The overarching idea of the Partners and Pals program is probably best explained by someone at Epic, but my layperson's framework for what it means is that partners are more established companies in this space. Press Ganey is a good example of a more established company that meets the criteria for being a partner, whereas Pals are earlier in their journeys. Abridge, as mature or well-financed as we might be, is a startup at the end of the day. We're not some company making billions of revenue every year. For us to build the trust and build a relationship with Epic to be one of the companies they've chosen to go deep with is certainly an incredible privilege that we don't take for granted.

Absolutely. We have thousands of clinicians using this across any number of specialties right now. Certainly, the bulk of the focus and demand has been in primary care, internal medicine, family medicine, and associated subspecialties. But there are any number of ENT clinicians, palliative care doctors, and other types of clinicians who are also using it on a daily basis. The company is still... if you look us up on LinkedIn, you'll probably see 30-some employees. That said, if you look us up in a few days, it might look very different. We're definitely in this mode right now in 2023. I've used this line over and over: we feel like we're hitting a different moment. There are budgets being dedicated to addressing this problem because we've just sort of pulled the rubber band as far as it can go before it snaps. What we truly believe, and what we're seeing, is that this kind of technology is never about technology; it's all about the people. We do believe that this is a technology that can actually help people deliver healthcare in a way that makes them feel more human.