This transcription is provided by artificial intelligence. We believe in technology but understand that even the smartest robots can sometimes get speech recognition wrong.

Executive Interview: Shrinking the Gap Between Idea and Reality with Nishith Khandwala

[00:00:00] This episode is brought to you by Bunkerhill Health. Healthcare AI often falls short, either surfacing problems without solving them or handling one use case in isolation. Bunker Hill's Care Bricks platform automates workflows end to end from addressing actionable findings to automating. Prior auth and any other workflow a health system requires in any specialty.

Care teams spend less time on manual tasks and patients receive faster, more reliable care. Learn more at this week. health.com/bunker Hill.

I'm Bill Russell, creator of this Week Health, where our mission is to transform healthcare, one connection at a time. This is an executive interview

quick powerful Conversations with Leaders Driving Change. So let's get started.

Bill Russell: All right. Today we have an executive interview Today I'm joined with Nish Khandwala, with Bunker Hill Health and Nish, welcome to the [00:01:00] show.

So nice to be here. Thanks for having me, bill.

I am looking forward to this conversation. I love the stories and the founder stories, founder and CEO.

So you're a grad student at Stanford. Take us from that moment till Bunker Hill becomes a reality.

Nish Khandwala: Yeah,

Bill Russell: absolutely.

Nish Khandwala: This was many years ago, back in, around 20 17, 20 16 when deep learning was just getting started. I was a computer science graduate student. I was working in a few AI labs and I was in the computer science role.

I was not in the healthcare side at all, but we were approached by. Many physicians from the School of Medicine with different ideas for how AI could help their practice. And I remember this one day we had the director of preventive cardiology from Stanford approach our lab, and he had this idea, which blew my mind, and I was like, oh, this is finally a great use case for ai for the clinical side of healthcare.

Basically the use case was as a preventive [00:02:00] cardiologist, he would see patients in his clinic. After they've had their heart attack or their first stroke, and his job is to prevent the future ones on many occasions, what he would find that as he would do chart review to see what led to the first heart attack, he would see that same patient.

Had come to Stanford Hospital four or five years ago for some entirely unrelated reason. Maybe they had come in for a lung cancer screening, maybe they were part of a car accident and had come to the emergency department, maybe they had a pneumonia and then visited a pulmonologist. And in the due course of care, they had a CT scan that was done on that scan.

He would notice that there was coronary calcium, which is a biomarker indicator of cardiac disease, and he would see that no one did anything about it. That patient literally needed to have a heart attack before they met with a cardiologist. And so his proposal to our lab was, could we build an AI algorithm that could comb through all the CT [00:03:00] scans that are being done at Stanford for some non-cardiovascular reason, and pick up those patients who are at high risk, get them in front of a cardiologist and potentially drive better care.

And we thought that was. the best thing since sliced bread. I'd be very ecstatic to work on it. We had clinical buy-in. We obviously knew that from a machine learning standpoint, this didn't seem that difficult. And so we worked with the cardiology group, the radiology group, to build that algorithm, and it worked.

And we published a paper. We made some incredibly bold claims about how this is gonna save hundreds of thousands of lives. Potentially impact lives of millions of patients and at the very least, revolutionized cardiovascular medicine. Two years passed by, none of those claims came true, not even close.

We had yet to impact a single patient, to touch a single patient, let alone save lives, and it was very embarrassing. When you make bold claims and things don't actually pan out, and [00:04:00] this happened a couple of times and we thought, this is why academia sucks. This is why people burn out.

This is why, computer science students like myself, would go to Facebook and optimize for ads as opposed to, work in healthcare. I was ready to do the same. I was like, okay, this isn't, this seems so weird that there's something that works. The CFO wants it. The cardiologist obviously wants it.

Why is this not getting adopted? I was ready to pack my bags, but then life had it another way. My dad had a heart attack. Thankfully he's fine now, but when they rushed him to the emergency room, they got a cardiac ct and. The cardiologist was telling me, here's what coronary calcium means.

Oh gosh.

Bill Russell: So you just apply your algorithm and he's been okay.

Nish Khandwala: Yeah. I was like, you don't need to dumb this down for me, because I spent a year working on an algorithm that did exactly this, and like, the coincidence was just, it was too close to home. It forced the question, why was that algorithm that we had worked on that worked [00:05:00] not being used?

At many hospitals across the world, why was it not even being used at Stanford? And I couldn't give a good answer. I needed to really spend some time digging up that. And we spend the next couple of years just trying to understand the question from the point that someone has an idea for how AI could impact a clinical problem or an operational problem.

How quickly can we bring that. Sort of idea to reality and widespread adoption. And that's why we spun out this company, which whose entire goal is to decrease that time to as little as possible. We could have started a company whose entire job was to commercialize that one algorithm that we had worked on, but we saw just how many researchers like myself across different labs were all facing this common problem of will you build something, it works.

And it doesn't go anywhere. Nothing happens. And so we spun out this company with the sole goal of helping people who have ideas for [00:06:00] how AI can help clinical and operational workflows. How do we get those ideas to reality as quickly as possible? And that's what we do at Bunker Hill.

Bill Russell: first of all, amazing story, amazing use case and the innovation usually comes outta frustration, right? So we get we get tired of a situation existing and saying, this is where we're gonna go. When I think of the problem set that you're looking at. You have to get access to clinical data, potentially some demographic data, maybe some genomic data.

bottom line is you need to get access to data in order to do what you're talking about and then apply those algorithms and then somehow get hooked into the workflows. Is that generally?

Nish Khandwala: Yeah. That problem used to be a lot more difficult. But then generative AI came and it became easier and easier to build those workflows.

Like today, for example, if you wanted to build a workflow, which for example, because of GLP ones, bariatric surgery, volumes have gone down significantly. And so you could imagine a bariatric surgeon saying [00:07:00] like, Hey, I wanna find patients who could really utilize bariatric surgery. If you, from a machine learning standpoint, you could take in patient's data.

You don't even need to build a new algorithm for this. You could take an off the shelf large language model, like open AI's, GPT oh three, for example, and basically ask questions about that patient's data to the large language model and figure out if a patient is a good candidate or not. And you could potentially implement a system that try to screen for those types of patients.

it has become easier and easier over time to build AI enabled workflows, but what has not? come easier is how quickly can you operationalize those workflows? How quickly can you take those algorithms or workflows that you've conceived of, that you have built and actually implement those in clinical practice?

So yeah, I mean, there's still many use cases for which you might want to go and build a specific. [00:08:00] Algorithm, but for a wide variety of low hanging fruit use cases, you don't even need to build an algorithm. The algorithm sort of already exists in the form of a large language model. The difficulty is actually operationalizing and creating a workflow and automatic system around it and that's what we are really trying to solve here.

Bill Russell: So you have a platform to essentially operationalize these breakthroughs. is it distinct from the EHR and how does it function with the EHR? So

Nish Khandwala: I think the EHR is a system of record that's very different from what we are trying to do, right? The EHR as we think of it is the way where it's the gold standard for truth.

That's what you look at when you're interested in learning about whether a patient. Has had a particular diagnosis, a particular progress, and so on and so forth. It's the database that stores that information and it's what the physicians interact with. It's what stores the information, gets new information added to it or deleted from it, it is a system of record. What [00:09:00] we are trying to build is a workflow layer on top of that where if you thought of an idea that basically. Followed the pattern of clinical reasoning followed by some kind of action. Then how quickly can you take what you have in your mind and create a workflow from it?

I'll take another example. Wegovy was recently FDA approved for fatty liver patients the patients with fatty liver disease. Previously if a patient was diagnosed with fatty liver disease, you as a physician didn't have much of a optionality than to just tell the patient, Hey, take better care of yourself.

Now. Suddenly there's a treatment that's available to you. So if you are someone who treats a patient for fatty liver disease, the first thing that comes to your mind is, oh, could I basically comb through all the patients I've seen in the past and see which of these patients would be eligible for wegovy?

Now today, if you wanted to do that, yours, you're a system that uses Epic, for example, you have to go to Slicer Dicer, which only, again, looks [00:10:00] at structured data, or you have to manually comb through a bunch of information, higher FTEs to do that, what if you could just use AI too basically create that flow of information, define that population of patients that is of interest to you, and then assign an automatic flow to them.

So these are the kinds of use cases that get us really excited. It's like, how quickly can you go from some kind of idea that you had that follows this pattern of clinical reasoning and action and implement that rapidly in your EHR. It doesn't matter. It's just we are not trying to replace an EHR by any means.

It's mostly a workflow layer on top of that.

Bill Russell: Because it's a layer on top of that, it gives you, I would imagine an awful lot of flexibility. the problems you can solve and the direction you can take it. I mean, it might be system of record, EHR data that you're pulling in, but you could be pulling in all sorts of data from various sources, I would imagine.

Nish Khandwala: Correct. Yeah. So we look at three sources of information. The first source of information is what's in the EHR about the patient. So this includes structured and unstructured data. [00:11:00] So we have an app on most common EHRs, as you can imagine. The second source of information is the internet. Could you look up things like research papers, clinical guidelines FDA labels clinical trials.gov, so on and so forth.

And then finally, the third source of information is any sort of specific rule set or SOPs that an institution might have for themselves. For example, cleveland Clinic has realized that aortic patients with moderate aortic stenosis could sometimes benefit from a valve replacement surgery.

Now, that's not technically in the guidelines, so they have a very specific sort of pathway for it for their group of patients. That's an internal SOP that they have. And that's something that we could also ingest and read and incorporate that. So we don't look at just the HR, it's EHR, the internet and any sort of internal documentation that a health system might have.

And then could you use that source of information to identify a population of interest and then assign a [00:12:00] flow in terms of downstream action for those group of patients in that population?

Bill Russell: who's the, Who's the champion within a health system? That you talk to and they go, we've gotta have this.

I mean, is it the researcher? Is it the specialist? Is it somebody is in charge of quality?

Nish Khandwala: That's a great question. I think we are as much of a. Tech innovation or a tech platform as much as an operational one. Now think of a chief operating officer or a chief medical officer, or now in many cases, a Chief AI officer at a health system.

They get inundated by people from within the system coming to them and saying, Hey, I think this is a problem for me. I would love to have additional headcount for this. Or in the case of the chief AI officer, it's like, Hey, I have this point. Solution could be on board that or procure that. They get bombarded with these sort of requests and this platform is intended for them.

It's like, Hey, could you adopt this platform as an enterprise strategy? Then enable different [00:13:00] departments, different groups of people within the health system to use that to solve their own problems. Now, I've only given shared examples thus far that are what I would call frontier use cases like the We Gobi example, or the coronary calcium on chest cts, the same platform, the same sort of paradigm of clinical reasoning plus action can also solve more cookie cutter problems.

Finding patients in a referral queue that need to be triaged up or helping with prior authorizations for specialty drugs or helping navigate patients with actionable findings. So we see in most cases that we've had success. It's typically someone at the enterprise level who says, Hey, this is going to be our AI strategy.

This is going to be our enterprise AI platform. When people have questions or when people have ideas for how to use ai, this can be where you start. This can be where you see with the impact of AI within your health system. So it's a platform for anything that follows the pattern of clinical [00:14:00] reasoning followed by some kind of action.

Bill Russell: So you're working with some impressive health systems. What kind of impact are they seeing in practice?

Nish Khandwala: So I'll share a couple of examples at University of Texas Medical Branch in Galveston, UTMB. We have over 15 use cases that are currently live now. That was just incredible. We started with one and there was just so much rapid adoption because when the Chief AI officer there.

Adopted this as their enterprise AI strategy. So many people from within the health system raised their hand and said, Hey, we have a use case for this. We have a use case for this. And it ranged all the way from cookie cutter use cases, like prior authorization for specialty drugs on the pharmacy side.

Two incredibly novel ideas around how to automate consults for patients with low hemoglobin, for example, so that the pathology team or the hematology team does not have to spend sort of, manual time doing that. At places like the Cleveland Clinic. We are helping [00:15:00] them identify patients who are at risk for cardiovascular disease.

And we have found patients in the orders of thousands that have a high risk for cardiac disease, but no existing care from a cardiologist. They have never seen one never seen a cardiologist before. They're not on the right doses of statins or they're not on statin entirely and could really benefit from seeing a cardiologist.

Like one very recently we had an instance where. police officer had a high calcium score around 400. We notified that police officer he had come in for some entirely unrelated reason. That patient came back to the hospital, saw a cardiologist, and turns out that this patient had chest pain, but they had previously dismissed it as just heartburn.

This is a, early fifties patient and turns out that when this patient was subject to some cardiac testing, they failed all that test and ended up having a triple vessel bypass surgery. Now those are instances where we are just like, wow, it's such a [00:16:00] simple use case, but the impact has been just massive.

Not to mention obviously the downstream ROI that a health system benefits from, in terms of more procedures. So on and so forth. Just the patient story itself there was incredible to observe

Bill Russell: The architecture sounds like a true platform. So you have the platform, then you have Care Bricks.

Describe for me what Care Bricks are. It just, it sounds modular like I Yes, it's a great name.

Nish Khandwala: That's the intention. So Care Bricks is how we implement these workflows. And as the name implies, as you astutely noted, it's a bunch of bricks put together like Lego bricks. At the core of it is our reasoning brick, which is a large language model accompanied by a library of FD acle tools that we have licensed from academic centers.

So we license AI algorithms that researchers have created at these health systems and bring them all to our platform. So the Reasoning Brick has access to the source of information that I'd mentioned, patient information, the internet, as well as hospital specific SOPs. And what [00:17:00] Care Brick allows you to do is use those information sources and the large language model, plus the algorithms, and create a workflow that ultimately translates in an automated action.

That automated action can be something like notifying a patient. Through text messaging through snail mail, through the EHR could be notifying, the physician could be writing back to the EHR by creating an order of some kind like a referral or even interacting with third party portals, like peer portals or pharmacy portals, for example.

And Care Bricks is the way to go from an idea to a automated workflow within a couple of hours. The buzzwordy way of saying this is this a platform to create AI agents? But I know AI agents nowadays mean everything under the sun. I don't start with that.

Bill Russell: yeah.

The question becomes are they autonomous AI agents? I think the last question I want to ask you is, as you scale this across more health systems, what excites you the most about where this is heading? I mean, do you feel like this is [00:18:00] going to address the problem you talked about earlier with your father and having the algorithm, and, I mean, and are you seeing that kind of progress?

Nish Khandwala: Yeah, I think as it becomes easier and easier to build. As AI becomes more powerful, I think it's gonna help solve that problem even more. So, our North Star is from the time you have an idea to it being live, could you shorten that timeframe as much as possible? So, I would love to live in a world where someone at a hospital has an idea.

They could either create that workflow within a couple of hours. Or borrow an existing workflow that a different health system has already created on the platform and just start using it. Wouldn't that be great? Like, I live in San Francisco and you just see the pace at which these AI companies are moving, and then you attend any healthcare conference and you're like, wow, this just feels like

things have slowed down substantially, even though compared to previous years, it still feels [00:19:00] faster, but just relatively to the AI world, I'm just looking for us to accelerate in a world where we can really go from idea to clinical adoption as rapidly as possible.

Bill Russell: Yeah. And really, I mean, what you guys are providing is the guardrails for it to go faster.

I sat with some CIOs and I was talking about all the. Things we're able to do in our business? Well, our business is not healthcare. It's media events, executive development and those kinds of things. I'm like, look, I can do this. I can do this. I'm using autonomous agents. I'm doing this, and I'm funneling all this information through.

It's almost like having a COO who sits inside of or data. Yeah. And is just giving me insights all day of, Hey have you considered this? And what about this tax strategy and stuff? And you're like, oh my gosh, it's like I just hired experts who are, giving me this feedback

and when I have that conversation with healthcare CIOs, they sort of look at me and go. I see what you see and I see the potential. we can't move at that page. Especially on the clinical side, like on the administrative side. They're like, okay, I could see it. [00:20:00] Help desk tickets and, those kinds of things.

But I mean, that is so scratching the surface of Yeah. What we're gonna be able to do with care.

Nish Khandwala: Yeah. People do share that. Hey, healthcare just doesn't move fast enough, and I haven't heard a first principal's reason for that yet. Yes, it's high stakes, but so is the finance industry.

They move pretty fast. Healthcare is obviously higher stakes. One could argue, but there's so much low hanging fruit like. Why would it be so challenging to conceive of a system that looked at every patient within the health system, looked at their records, scanned those records every day to see is there an untreated finding?

Is there an unresolved risk factor and the patient should be called back in? That doesn't sound very sci-fi from an AI perspective. And I see no reason why something like this could be considered as super high stakes and can't be something that we can't implement fast enough. I think you really have to [00:21:00] equip the health system with the modular.

Again, obviously I'm biased here with the modular and a platform that you can rapidly implement these types of workflows and use cases.

Bill Russell: I wanna thank you for coming on. I want to thank you for sharing your story. Where can people get more information about Bunker Hill And start a conversation with you.

Nish Khandwala: Yeah. You can go to bunker Hill health.com. You'll find more information about about us there. You can also email me my, it's my first name, nsh@bunkerhillhealth.com. Would love to hear from your audience,

Bill Russell: NSH. I'd be remiss if I didn't ask bunker Hill.

Nish Khandwala: We are not a Boston company, even though people might think or find that as a reference, there was a TV show in 20 16 20 17, which I would not recommend you watch. Uh, It got canceled midway through the first season. That's how lame it was. The TV show was called Pure Genius. While it was a lame TV show it was centered around a hospital called Bunker Hill. The research that was done in the morning was used in clinical practice in the afternoon.

That was the translation speed [00:22:00] that we were looking for. And so we called our company Bunker Hill after that hospital. There you go.

Bill Russell: And it is very memorable, so that's fair. I appreciate it. N looking forward to keeping this conversation going as I'd love to see the progress you guys make over the next year.

Thank

Nish Khandwala: you,

Thanks for joining us for this executive interview with me, bill Russell. Every healthcare leader needs a community they can lean on and learn from. Subscribe at this week, health.com/subscribe and share this conversation with your team. Together we're transforming healthcare.

Thanks for listening. That's all for now.