HIStalk Interviews Manny Krakaris, CEO, Augmedix

Manny Krakaris, MBA is president and CEO of Augmedix. 

Tell me about yourself and the company.

I spent the first 12 years of my career in banking. I then made the transition to industry because I wanted to get more of a hands-on experience in business. I’ve been a serial CEO, COO, and CFO of a variety of technology companies in semiconductors, solar, enterprise software, and SaaS. I came to Augmedix via the board. I had previously worked with two of the board members who were representatives of VC firms whose portfolio companies I had worked for. I had just sold my previous company.

I was intrigued by the opportunity. My doctor was a customer well before I met with the board, and she just loved the service. She had told me in no uncertain terms that this had changed her life. I found that intriguing and compelling. When I did a little bit more homework, I came to realize that this was a pervasive problem. Having been new to healthcare, I had no idea that this was a problem in the first place, but it was pervasive and huge. I felt that this is an area where I could contribute, given my background, to help bring a little bit more efficiency into the healthcare sector.

How did the pending acquisition of Augmedix by Commure happen?

We share a common customer, HCA. We provide different services to HCA. HCA gently encouraged us to start talking about how we might be able to stitch together a more comprehensive suite of solutions that addresses a wider swath of the patient journey when it comes to healthcare. The more we talked, the more interesting it became. It ultimately culminated in Commure making the offer to acquire us last month.

Commure had announced its own free ambient scribing solution three months before the acquisition announcement. How does that fit with its strategy?

They did have that offer out there and they still do. The idea is that by offering a platform with a whole suite of products, you can bundle things so that a specific offering can be made available at a seemingly low price or even free in some cases. I don’t think Commure is the first to come up with that concept. Microsoft has been pursuing that strategy for forever, it seems, and quite successfully. In our case, they have an ambient AI scribe product, but it caters to a different care setting than the ones that we focus on, so they are quite complementary. Down the road, will we share similar back ends? Probably, but time will tell.

How do you differentiate your product from the several competing ones?

At last count, I think there are 42 companies that are purporting to be able to generate a note using large language models without any human intervention. The reality is that you can create a draft medical note from the use of automatic speech recognition to convert an audio recording into a transcript, and then large language models take that transcript and convert that into a structured medical note. But the structured medical note that comes out of the back end is a rough draft that requires human intervention to complete it, to edit it, and to make sure that there are no hallucinations in it. The state of the technology is not perfect yet.

What differentiates us from the vast majority of those companies is that we approach this problem organically. We pioneered the whole concept of ambient medical documentation 11 years ago, when no one had ever heard of it. It was revolutionary to basically tell the industry, look, we can repurpose the conversation that occurs between a doctor and a patient and use that as the primary input source to create a medical note. What technology has helped us do in the last couple of years is automate that last step using large language models. If you simply try to modify the technology to this particular use case, you’re won’t get good quality output.

We understand clinician workflows better than pretty much anybody, with a possible exception of one company. We also understand the differences of clinicians’ needs based on care setting, specialty, and the complexity of the encounter. We incorporate that into the portfolio of solutions that we offer today. One size fits all does not work in healthcare.

How important is being able to complete the note quickly, ideally just before the visit ends?

Obviously speed is important. You don’t want to have your customer waiting for minutes or hours for their medical note, because they need to move on to the next patient. For the self -serve products that are fully AI capable, you want to be able to get that draft note to the clinician within a half a minute or so.  Several players have been able to reach that milestone.

Will low switching costs encourage customers to change vendors?

Switching costs with software of this type, which is downloadable application from Google Marketplace or the App Store, are going to become less significant than they were in the past. It all depends on how deeply integrated the application becomes in the clinician’s workflow.

For independent practices, the degree to which the application is integrated in the workflow is pretty low. I would imagine that for that segment of the market, switching costs are going to be insignificant. But for the enterprise, there are significant points of integration with the EMRs, RCM, and patient intake that would make switching costs much more prohibitive for the incumbent to have a greater moat established around their business.

How does the ability to take action from the user’s voice commands overlap with ambient documentation?

They’re pretty much the same thing. Ambient is all about voice. It’s taking the voice recording between a doctor and a patient and using that to generate a medical note. Voice commands, in terms of requesting data from different parts of a healthcare system, are just an extension of the ambient technology. I think that is going to become more and more prevalent. It’s already pretty pervasive in some healthcare systems. I don’t see that reversing. That’s a big efficiency gain for the healthcare industry.

Ambient documentation seems to have higher physician acceptance than most technologies. What is the rationale of those who choose not to use it?

I think we have to stratify the market, which is true of any industry, not just healthcare. When you introduce new technology, you’re going to have some enthusiastic early adopters who want to see change and want to help shape that change. That’s what we’re seeing today in healthcare. The preponderance of users of self-serve AI tools today, whether they are our customers or customers of our competitors, are for the most part early adopters. They are willing to put up with some imperfections in the technology and provide input to make that technology better.

For mass adoption to occur, you need to remove any kind of friction points or imperfections in the technology. I think we’re going to see more and more of that towards the end of this year and certainly in 2025 as the technology matures a little more. It’s not quite there yet, but it’s getting there.

Is it hard to make ambient documentation work as well for specialists and nurses as it does for primary care physicians?

The technology is only as good as the input that you put into training it when it comes to large language models. GPT-4 is a very powerful general purpose tool. If you prompt it with a general question such as “create a structured medical note in these four different segments based on this transcript,” it will do that for you. It will be OK, but not great.

However, if you start asking it more refined questions — for example, if you do what we do with proprietary models that identify in the transcript the key elements of that transcript that you believe are relevant to the medical note — and then you ask it specific questions for each one of those elements, you narrow the variables that the LLM has to deal with to generate a response. The fewer variables you ask the model to work with, the more accurate your output is going to be. That’s what we do. We ask very specific questions of each of the key medical elements that we identify in the transcript in order to optimize accuracy.

Beyond reducing after-work chart completion, does ambient documentation reduce the cognitive load of physicians who otherwise would need to listen and type at the same time or try to recall parts of the conversation to create documentation after the fact?

Yes. We conducted a study with one of our largest customers. It was a pretty broad study that included primary care physicians and a variety of specialists, well over 100 clinicians whom we studied over a year.

We discovered that for primary care physicians, the biggest source of improvement in their WRVUs — their work relative value units, which is a standard measure of performance of a physician — did not come from increased patient throughput. Rather it came from higher capture rates, which then resulted in higher reimbursement.

It’s not intuitive at first, but if clinicians have to try to remember everything that they did during an encounter when they subsequently do the medical note, several things may slip through the cracks. That is, in fact, what has happened, in our study at least. Those slippages, those things that were omitted, represented about 80% of the lift, and the lift was significant. You can add value beyond increasing patient throughput or reducing pajama time.

What is the near-term future for using AI in healthcare?

AI has the capacity to learn quickly. The rate at which it learns really depends on the rate at which you can feed it relevant data. It will be incumbent on healthcare systems to ensure that the data that their vendors are using to train their models is representative of the patient population of that particular healthcare enterprise. It’s not good enough, and in fact is  counterproductive in many ways, to use generalized data from the general population. If you’re trying to cater to a regional healthcare network that caters mostly to foreign-speaking people or people of a certain ethnicity who are not represented equally within the general population, that will skew how the model interprets certain information. It’s important to tailor the data that you use to train your models to the patient population that your customers are serving.

Second, as you train the models, you can actually help the model mimic the preferences of the individual clinician, looking at what the clinician does from an editing perspective after the draft note is delivered to the clinician. Take those edits that the clinician makes to what the technology generated and put that back into your training data. That will generate a note that better reflects the preferences and stylistic preferences of the doctor. That’s going to be welcomed by many doctors, because they have their own unique ways of documenting their interactions with patients. AI has the ability to to learn from that as long as we can get that that feedback and incorporate that into the training models.

What does the post-acquisition future of the company look like?

This is my first foray in healthcare, so I come into this with a naive perspective, but if you follow the patient journey, it has many steps. Each one seems to be provided by a different entity that is providing a very specific task. If you look at it holistically, to go from patient intake to final reimbursement, there are way too many disjointed steps in between.

What I think the healthcare industry could benefit from greatly, which is lacking so far, is compressing as many of those steps as possible by integrating them on a singular platform that seamlessly transfers information from one functional area to another to another to avoid what happens today, which is a lot of manual intervention to clean up imperfect input from the preceding functional step in that journey. That introduces a lot of cost in the system. That’s something that the healthcare industry really can no longer afford to do. Commure’s vision is to be the first in the industry to be able to do that. I think we play a central role in that strategy.

The healthcare industry is intriguing. It’s massive. There are a lot of challenges in front of us, but I think the people that run the big hospital systems, healthcare networks, and IDNs, are of the mindset today that doing the same thing is not going to yield the kind of results they need to generate in order to be able to continue to deliver healthcare to a growing and aging patient population. They are a lot more willing today than they were six years ago, when I got into this industry, to explore these new opportunities and new technologies. I find that very encouraging.