AI Bringing Care to Remote Areas

In this episode of "Full Tech Ahead," host Amanda Razani interviews Dr. Jason Corso, Toyota Professor of AI at the University of Michigan and Co-Founder of Voxel51. They discuss Voxel51’s role as a developer tool software company for physical and visual AI, which has achieved over 4 million downloads. 

The core of the conversation focuses on Vigil, an innovative healthcare AI project led by Dr. Corso and funded by ARPA-H’s Paradigm program. Vigil tackles the critical shortage of specialists and brick-and-mortar hospitals in rural America by equipping mobile medical units (clinics on wheels) with physically grounded AI. 

Instead of replacing clinicians, Vigil acts as an advanced co-pilot, using computer vision and on-the-fly micro-guidance to upskill generalist healthcare workers (like registered nurses or EMTs) to perform complex procedures, such as cardiac ultrasound diagnostics, directly in remote communities.

Key Quotes

• "Voxel51 is indeed a dev tool software company for AI that supports the developer... in the spaces of physical AI and visual AI."
• "I don't think AI is here to replace humans... I just believe that we are as technologists in AI, we are building tools that will augment humans."
• "We have this notion of a triangle of trust where the healthcare worker is trusting Vigil to help him or her, and the patient is trusting the healthcare worker, and then tacitly, the patient is trusting Vigil."
• "In the healthcare, in the visual domain, we can't hallucinate, first of all... We're really trying to get toward those guaranteeable guardrails."

Takeaways

• Upskilling the Generalist Workforce: AI can dramatically expand healthcare access without needing to "clone specialists." By equipping existing local nurses or EMTs with AI-guided tools, they can perform specialized tasks—like capturing precise cardiac ultrasound imagery—that normally require years of dedicated training.
• The "Triangle of Trust": Successful AI deployment in healthcare relies heavily on the bedside manner and human connection. The patient trusts the clinician, the clinician trusts the AI, and the patient tacitly trusts the AI. Maintaining this human-centered relationship is crucial.
• Guaranteeable Model Guardrails: Unlike conversational LLMs that are prone to hallucination and rely on post-hoc prompt filters, critical visual AI systems in healthcare require deeply grounded, mathematical, and theoretical guardrails that prevent errors before they happen to ensure patient safety.
• Augmentation over Replacement: The future of advanced technology, including robotics (like actuated robotic arms in mobile clinics), is to augment human capabilities. AI provides an extra set of un-blinded eyes and precise micron-level assistance, allowing human workers to perform their jobs faster, better, and more equitably.

Find Amanda Razani on LinkedIn.  https://www.linkedin.com/in/amanda-razani-990a7233/

Follow the FTA LinkedIn Page: https://www.linkedin.com/company/full-tech-ahead/

Visit the FTA website: https://fulltechahead.com/

Check out the Substack Channel: https://fulltechahead.substack.com/

Transcript

SPEAKER_00 0:19

Hello and welcome to Full Tech Ahead. I'm Amanda Rizzani, and with me today I'm excited to have Jason Corso. He is a Toyota professor of AI at the University of Michigan and co-founder of Voxel 51, a company focused on visual AI. He's also leading an exciting AI project with ARPA H called Vigil, which is exploring how AI can help extend healthcare access in remote communities by supporting clinicians rather than replacing them. So, Jason, welcome to the show. Can you share a little bit about Voxel 51, first of all?

SPEAKER_01 0:57

Absolutely. Thanks, Amanda. It's great to be here with you. So Voxel 51 is indeed a dev tool software company for AI that supports the developer, like the machine learning scientist, machine learning engineer, data annotator, data curator in the spaces of physical AI and visual AI. Specifically, we make a software tool called 51, which supports the full stack of the workflows from managing, curating, annotating data through training, evaluating, and deploying the models that are trained from that data. We work in various verticals like automotive, healthcare, physical security, product support, generally are supported by a uh we're about 60 people now, actually, which is great for me. You know, it's it's been massive growth in the last five, five or so years. Uh we have about four million downloads of the open source software. This is like a single user instance of 51. And we uh support both on-prem and air gap solutions uh when data security is really important to our customers.

SPEAKER_00 2:01

Awesome. Well, I want to talk more about this vigil. What problem and role healthcare were you trying to solve when you created this?

SPEAKER_01 2:11

Awesome. Yeah. So right at the University of Michigan, I have I had been getting more into healthcare uh in the last three or four years and was delighted when this program from ARPH, the Paradigm Program, came online. So um not being a physician, you know, this is I'm an engineer, right, by training. So having to learn about physical physician or medical-oriented problems, healthcare problems in the US really is enlightening to me. Um so Vigil really tackles a um a key problem in rural healthcare in the U.S. So generally, there's a lack of specialist, a lack of expert-trained physicians, so MDs or DOs that operate in rural America. And to make that problem even worse, there is an increasing shortage of physical brick and mortar locations where rural Americans can go nearby their homes to get healthcare, i.e., hospitals are closing, even just the clinics are closing. There's one study I'm aware of that was based in North Carolina that measured the distance a North Carolina citizen had to travel, the average distance, to get to their nearest brick and mortar healthcare facility between the years of 2012 and 2018 increased eightfold. So that means someone was driving 15 minutes to get to their nearest hospital. Now they may have to drive two hours to get, or more, frankly, because this is an average. Um, so so ultimately, vigil uh is there to well, obviously, we can't just, you know, clone physicians, we can't build brick and mortar hospitals, right? That would have been done by other institutions already. So what we are trying to do is, well, first of all, we make the observation that although there are few specialists, few expert-trained MDs and DOs, there are actually quite a quite a large number of registered nurses or physic physicians assistants or even EMTs that are in those regions. And although they may not be specialists in any one particular area, they are generally trained with medical knowledge. And so the idea is if we can equip them with kind of like a hospital on wheels or like a clinic on wheels, um, and some physically grounded AI, right? So kind of like a Hay Siri mindset, right? But uh but a Hay Siri, which we actually call Hay Vigil, right, with with a vigil that actually has cameras and eyeballs and ears inside of this mobile clinic, so that when the generalist is doing their work with the patient, uh they're actually able to be guided on the fly by this AI system that we call vigil. Um clearly the this this generally trained or a generalist um healthcare worker has the final say, right? If they're ever not confident with what AI guidance they're getting, they can always phone home, you know, phone back to the hospital, may elongate the visit. However, the idea is that we'll be able to deliver significantly better, uh richer, and more robust healthcare in rural America by building these mobile vans, staffing them with one or two of these generalist healthcare workers, and then giving them a road, you know, affiliating with a with a rural hospital, obviously, right? Because we have to have, you know, we're not gonna have patients of our own, but affiliating with a rural hospital, and then they'll have a day in which they can treat maybe 12, 16, 20 different patients who otherwise would have had to take days off, full days off of work or what have you.

SPEAKER_00 5:29

That is fantastic. Well, I can say definitely, I live in the heart of Texas, and there's a lot of rural communities all over that are all reliant on a hospital, like you said, uh hour, two hours away. So that is very helpful. So why did you take an approach that helps clinicians instead of trying to just replace them with AI?

SPEAKER_01 5:53

Well, I think this goes back to my general philosophy. Um, you know, I I've been in the field for about 25 years now. I just I just believe that we are, as technologists in AI, we are building tools that will augment humans, right? Humans are the center of our experience. We are humans, right? In some sense. And I have this saying now that I've been working with a colleague of mine, for all humans, for all life, right? Like humans are the center of our world. And I don't think AI is here to replace humans. We probably could have another conversation about whether or not like white-collar jobs are going to be lost to AIs. I just don't believe so. Frankly, I think it's gonna be a boon for white-collar jobs because they'll have better tooling around what can be done. Uh, in in this particular healthcare space, specifically, uh, you know, we can't, we could not go out and do deliver vigil in practice today because of various state-level um like laws, essentially, like liability laws and you know, very various safety laws that are that are put in place for great reasons to protect patients. Um and so we we just frankly believe the right way today, tomorrow, in the next five years to deliver this increased level of rural healthcare is to just augment the humans that are uh in in the vehicle that will be in the vehicle to do the work that they want to do, that they know how to do. You know, it's one key reason there is that, right? Like just tooling for these human AI teams. That's been the heart of my career in some sense. Um, you know, another aspect though is that, especially in healthcare, there is a significant value to the bond, that the trust and the bond between the healthcare worker and the patient. And vigil is there not to do in any way uh violate or corrupt that bond. In fact, we we we have this notion of a triangle of trust where the vigil worker, right, the healthcare worker is trusting vigil to help him or her, and the patient is trusting the healthcare worker, and then kind of tacitly the patient is hence trusting vigil. And if that triangle doesn't get built and get supported, then vigil will not be a success.

SPEAKER_00 7:49

Yeah, I agree that personal connection and that bedside manner are so critical. I have actually been in positions on a few occasions. Um, my mom, for example, uh lives in a rural town where um there's not as many eye doctors available, and they roll up a screen. Um uh, and even that is a real person on the screen, but it's still a little odd when you're just talking to a screen, you know, instead of going to, you know, you go to the eye clinic to see the eye doctor, but you just get a screen rolled up. So yeah, I understand.

SPEAKER_01 8:27

Uh I understand. In fact, in fact, um I I believe there are on the order of 30,000 plus mobile health units in practice today. So your your kids are right. So she's going somewhere and then they're rolling up a screen where where someone's being video telecomed in. But the notion of a mobile clinic is not novel to Vigil or to this ARPA paradigm program. I think what's what's actually what's really novel, I believe, is that there will be a human operating in the vehicle and there'll be upskilled, right? This generalist learning how to do things on the fly or not even just learning, but like being being reminded how to do things, right? So one for example, one one one uh clinical service we we're we're studying right now is cardiac diagnostic health through ultrasound. You know, typically an ultrasound operator need must take, I believe it's about two to two years, maybe three years of training to actually like be uh credentials in to perform an ultrasound for cardiac health. You know, it's it's great. And I'm very grateful that we have humans who go and do that work. And still we can't, but there's not enough of them, frankly, not only in rural places, but even in in like urban, semi-urban and urban places. So um, in this case, instead, someone, you know, an RN who might have touched an ultrasound three times can be upscaled on the fly to go and do the same level of care, to get the same quality images that either the computer will read automatically or will just be stored, and then a radiologist would read post hoc, um, which is not that different than going to a brick and mortar uh facility. You know, the the program manager, uh Dr. Bon Ku from RPAH, uh, the program manager of Paradigm, he was visiting campus last week actually for a site visit. And uh in in his morning presentation, uh made the made the comment that he believes on the order of 80% of all medical care does not need to be done in a brick and mortar facility. It could be done in the patient's home or in in the Walmart parking lot or in one of these mobile clinics. And I think realizing that healthcare of the future, or like laying the ground, the stonework, the groundwork for the healthcare of the future, you know, it's just it's frankly exciting to be a part of that.

SPEAKER_00 10:36

Yeah, it is. I mean, you think back to when you had those home office visits from your doctor. I mean, if you're really sick and ill, trying to get out and go somewhere, it's difficult. Not to mention the comfort of being in your home. So there's a lot of positives for sure. Well, and as you said, kind of breaking down, you know, you we're lacking in these healthcare areas. So kind of breaking down the time barriers, uh, making it easier to get people out there to help.

SPEAKER_01 11:05

Absolutely. Um, indeed, I I remember, in fact, when my spouse was about to deliver our first child, the chiropractor came for a home visit. Uh, and that was uh that was a big help, in fact. Yeah. But these home visits can it just changes the dynamic, really, right? So not not that there's uh well, I mean, the thing is it could not not like levels of playing field, but uh really just the the notion that all humans are valued and all humans are equal. I I think it it becomes more realizable when when we have when we have the ability to deliver better better health care.

SPEAKER_00 11:39

Absolutely. Well, you know, AI can sometimes be a hot button topic. So when it comes to AI and healthcare, what are some things that people get wrong or some misconceptions?

SPEAKER_01 11:51

Oh, it's a good it's a great question, right? So um, well, first of all, indeed, there is a there's a huge marketing machine behind AI that broadly, you know, both in healthcare and outside of healthcare, that unfortunately I think is doing a little bit of a disservice to actual AI technologists. Uh, you know, in the sense that uh there are increasing claims about when AI will solve problem X or we'll take job Y, or we'll have self-driving cars, you know, 10 years ago, and we still kind of don't have them, right? So like there's just this challenge where the message and the hype around AI is not actually backed out by robust engineering-based testing and delivery. Um, you know, in in the healthcare domain, you know, vigil is is um it treats the problem a little differently than we might think. Um, so the the idea in Vigil is to first of all do no harm, kind of like a you know traditional healthcare mindset, but that has different layers of impact, right? So some generalists maybe they're actually rather well trained in in ultrasound, just as an example, right? And some generalists are not trained in ultrasound. So if vigil gave the same level of guidance to both of those types of individuals, it would be a problem, right? They'd be distracting the more experienced individual. And so what when we are modeling guidance, uh we're we're actually using very contemporary state-of-the-art artificial intelligence methods. You know, we have our own healthcare world model that sits behind the technology, and then an agent that is modeling not only what is happening in the environment, but also what should be happening in the environment in the context of the knowledge about the skill level of the uh operator, of the generalist. And so if that agent decides, like that in some sense, the goal of that agent is to decide whether or not to provide guidance or just let what is good good, right? Like, you know, we have these multiple screens happening in the background that just kind of walk the generalist through the steps of the procedure that they should that they should be doing. Those the generalists can look at or not, right? Those are not intrusive. But at some point, for example, when you're holding an ultrasound probe near the heart, and you need to get like what's called an apical forechamber view to get an ejection fracture, how healthy, how much blood is getting pumped out of the left ventricle. Sometimes you have to make these micro adjustments, you know, it a fraction of a degree. And you really just build this sense over time. Um, so Vigil can opt to give different levels of training or guidance at that point. Uh, you know, it can render, for example, through a projector, yeah. We actually create this kind of shared perceptual space through a projecting the interface onto the patient's body. And so it can render different angles, different arrows. Uh, it renders like a ghost probe. So you can, you know, map the real probe to the ghost probe just to really have different ways of interacting with the patient. And I think that that type of uh modulated interaction through AI uh is is really is frankly really novel uh because we're not you know, you as you probably have heard, like these general LLM type models, they they're prone to hallucinate. And in and right, and so and and if you tell them they're hallucinating, they'll quickly believe you, right? There's actually not much deep modeling about the veracity of the conversation. Whereas in the healthcare, in this in the vigil domain, we can't hallucinate, first of all. So we really have to be grounded into what's allowable. Um, but we also have to model the fallibility of the generalist, the the likelihood that they may or may not make a mistake. Uh, and if they are making a mistake, we need to both anticipate it and then prevent it, especially if there's any any potential harm or whatever. So modeling that like safety in the interaction is super critical. Uh, and it's not something we've really seen enough of yet in the general AI space. I think that's one thing we're collectively getting wrong, right? Like um, we we don't really have appropriate uh in-model guardrails. From my understanding, most of the guardrails that are in the systems that we are all many of us are using on a regular basis, are post or pre-hoc or post hoc, right? They're they're guardrails on what could be set in the prompt or what could be outputted by the model to the to the user. Um, and they're not sort of guaranteeable in any notion of like theoretical or mathematical sense. And we are in vigil, we're really trying to get toward those guaranteeable guardrails so that um uh we you know we we we can really deploy this for for the for the greater good. Uh, you know, we haven't we have not begun an FDA process yet. We're only 18 months into a five-year project, right? So so we are developing this, but but um we are thinking already, you know, we've had a couple of conversations about that, and like how do we actually go and guarantee visual safety in the practice? Yeah.

SPEAKER_00 16:42

Absolutely. So it still has a ways to go. And to that point of the hype around AI and how fast it's it's building, um, you know, you hear a lot of things about, oh, soon there will be AI robot, physical robot doctors and nurses. What are your thoughts on that? And would there be a time where that's even needed? What are your thoughts?

SPEAKER_01 17:04

Yeah, I think it's a good, it's a good question because we all, you know, we've watched the sci-fi movies, we we listen to the media and so on, right? So I mean, I think in order for there to be any notion of an AI, you know, robot doctor, we need to have uh we need a few pieces in place. We already need to have a like a platform, like a physical healthcare platform that is robust and deployable and has and has certain fail-safes in it that could always um fit fall back to the expert human to make a decision. I think that ultimately to me that's going to be required in healthcare. Um, but we also need a significant amount of um social um education, social awareness, and kind of time, I think, right? Like time heals all wounds. Like I think time is needed to go and work through these changes. These are revolutionary. The potential for that type of technology is revolutionary. And so we can't just throw it out the user, throw it at throw it at all all Americans, throw it at all the world citizens, and like expect them to adopt it. And so I think when, you know, even even in the Vigil project, right now all the all the generalists are intended to be humans, and we expect that to be the case. That doesn't mean we aren't envisioning a future in which uh we could have a um a robot arm in the vehicle that's maybe mounted on the ceiling or mounted underneath the table or the chair where the the human the patient is sitting, that is getting the best optimal view so that vigil can make the best decision. Right? That's not a robot doctor, but that is a you know actuated robotic arm in the vehicle uh that will help the human make the better decision. In some sense, you know, you know, like humans can only see out of their eyes, we're only looking one direction, but with a robotic uh helper, we can have a whole different angle, maybe from the side view, for example, of you know, to round out and build up a more rich perception. And this is, I think this is the pathway toward a future in which there's just generally better healthcare provided, whether or not it's from a human or a future robot. You know, and I I mean I've I experienced, I was lucky enough to do my graduate work at Johns Hopkins, uh, because this is you know, this notion of a robot augmenting or helping a human in some way to do healthcare is not in any way new. Uh, you know, at Johns Hopkins, in the they had an ERC for computer integrated surgical systems and technology. Um, and Dr. Um Russ Taylor was the PI of that, of the NSF funded center, and they had this home-built robot called the steady hand robot. You know, and it could take movements of the operator, of the human surgeon, and um reduce them down to like micron level. Um, I'm I'm speculating a little bit, I think it was micron level, at least nano level, so very, very tiny level motions, so that you know, an eye surgeon, to use the ophthalmological ophthalmological example, could operate, you know, in a very sensitive part of the human body. Um, and and you know, this was like, I mean, I'm getting old now. This was like 25 years ago, right? To see this. So we're on this pathway, I think, of technology to augment humans doing good for you know in healthcare with with the with with new technology.

SPEAKER_00 20:16

Yeah, absolutely. That's the exciting part. Well, if there was one key takeaway you could leave our audience with today, what would that be?

SPEAKER_01 20:24

Well, I think the the key takeaway for me would always be that AI is here to augment humans and to you know to build these teams in which uh what we collectively can achieve will be magnified through the help of new technology, once delivered safely and responsibly.

SPEAKER_00 20:41

All right. Yes, indeed. Well, thank you so much for coming on the show and sharing your insights and all the new technology that's out there in the future.

SPEAKER_01 20:51

Thanks, Amanda. This was fantastic. Glad to talk about it.

SPEAKER_00 20:54

And thank you to our audience. If you have any questions or comments, leave those below and I'll try to answer them as soon as possible. And have a wonderful day.