What has historically taken place in an encounter with me is now able to happen at the patient’s home at the cadence of when it’s clinically indicated, not at the cadence of the appointment with me
Sharif Vakili, M.D., M.B.A., M.S., a primary care physician and clinical assistant professor at Stanford Medicine, is CEO of startup UpDoc, which has announced FDA clearance for what the company calls the first agentic clinical AI platform designed to support doctors. He recently spoke with Healthcare Innovation about the platform, which is being deployed at Cleveland Clinic, Allegheny Health Network, and UCSF Health.
UpDoc secured $18 million in an oversubscribed seed financing round from investors including the American Diabetes Association and Mayo Clinic
Healthcare Innovation: You are working on a new clinical AI-based model of care called remote patient intervention (RPI). Could you describe the most obvious use cases for this? Is it the primary care clinician with chronic care patients who would like to have more frequent interactions with them — and the AI can help with that?
Vakili:That’s right. It’s basically delivering continuous care in the interval between clinical visits in the chronic disease management setting, which tends to be primary care, but not exclusively. It’s also historically where remote patient monitoring had been, but now you can close the loop on it. Oftentimes what happens — and this has happened to me as a primary care physician — I’ll have patients who enrolled in these programs and I don’t even know about it, and it’s not just RPM programs, it’s any of these tech-enabled services or employer-facing programs. But now while delivering chronic care management, I’m able to close the loop on care delivery for my patients, and I do that in a way that is very integrated in what I’m doing day to day with the patient, as opposed to with variants of something happening out there in the ether.
HCI: Does this allow an AI agent to do some of the tasks that the clinician was having to do before or maybe wasn’t getting done at all?
Vakili:Precisely, and that’s actually what we call clinical AI. Something crosses the threshold of clinical AI when the AI agent is doing something that would have historically taken place in an encounter with the licensed clinician
HCI: Let’s talk about a few examples of that. Could it be adjusting medications or doing a lab order?
Vakili:An example is starting a medication or stopping a medication, adjusting the dose of the medication, or delivering a follow-up test. What has historically taken place in an encounter with me is now able to happen at the patient’s home at the cadence of when it’s clinically indicated, not at the cadence of the appointment with me
HCI: But is it done totally autonomously by the agent? Is there a scheduled meeting with the patient, and the agent is able to see that because of the lab values recorded, the patient’s medication dosage should be changed?
Vakili:The physician is very much involved in this. The patient is talking to the agent, and the agent adjusts the dose of the medication, but I’ve permitted it to do that ahead of time. I’ve neatly bounded what it’s allowed to do, and that’s very important. There are no surprises, and it’s not allowed to step outside the borders of what I’ve authorized for that patient
HCI: Could you give an anecdotal example of a patient with a chronic condition who’s seen under this scenario, and a few of the ways it’s different?
Vakili: I’ll give a traditional example, and then I’ll give onr in a world with RPI. We define RPI as when remote patient monitoring [RPM] data is acted on by a clinical AI to deliver care under the physician’s direction
Let’s say we have a patient with diabetes. That patient is being managed with insulin to manage their blood sugars. Probably I’d start them on some amount of insulin based on their weight, and I’d say, “Please check your blood sugars at home and bring those values in to me, and I’ll see you at the next appointment.” Usually, that next appointment might be three months down the line, when availability is, and throughout that entire interval, the patient is taking their blood sugars, and taking their medicine, but what I’m relying on them to do is to accurately collect that, bring their numbers back into clinic with me, and I can look at them, and then I’ll make an adjustment, and you rinse and repeat. This can take a long time. We see from studies that this can take more than a year to get controlled. That’s the traditional regime.
In this new world, when I feel it’s appropriate for my patients, I can deploy clinical AI to support me along with care. I’m still diagnosing and managing the patient. I’m doing this with the consent of the patient, and there is shared decision-making. Patient involvement is very important when you manage chronic disease. But now instead of prescribing the insulin, I’m prescribing a treatment plan: Here’s what I want to happen to the patient over the next six to 12 months, and then the clinical AI implements that treatment plan for me. It calls the patient, messages the patient, and does that plan, so the patient doesn’t have to wait three months until the appointment to have their medications adjusted. The agent will do that…
HCI: Because it has the clinical guidelines built in, so it knows the adjustment you would make?
Vakili:Correct. I tell it what to do. By the way, I don’t need to worry about the patient forgetting to do their homework and bringing that in to a visit. The care is broken down into little bite-sized pieces, and it happens at exactly the moment that’s clinically indicated. It’s almost like having a fastidious concierge provider checking in with you and closing the loop on that RPM data, right? You’ve got that CGM [continuous glucose monitoring] data, and now the clinical AI is following up to actually make the intervention happen. So by the time I next see that patient in clinic, in my next available slot in three months, the patient may already have had those meds dose-adjusted and all the follow-up tests collected. It’s a re-imagination of clinical care.
HCI: When your team did a clinical trial of this at Stanford, were there certain clinical outcome or financial improvements or measures of physician burnout or workload you were tracking?
Vakili:This was six years ago and a research system that the co-founders of UpDoc worked on, but it wasn’t UpDoc at that time — but it was this care delivery model. It was a randomized controlled trial. We had patients who were managed through the traditional method and patients managed with RPI. These were patients with uncontrolled diabetes. These patients were followed for two months, so it was a short time frame. But by the end of the trial, of the patients receiving the standard of care, about 25% of the patients got glycemic control, which is not too bad for this kind of cohort. But with RPI, 81% of the patients got control. There were five-fold more prescription adjustments in that arm of the study. Literally, that arm got five times more care delivered, right?
If you get five times more care delivery, you’re going to end up with better outcomes. It’s not particularly mysterious. When we look at chronic disease outcomes, we know the concept of therapeutic inertia is important. Making sure patients don’t slip through the cracks and they continue to build on what they’re doing, and continue that therapeutic momentum — there’s good literature on this. The second thing is this helps scale labor significantly. We had five-fold the amount of care delivery in one arm compared to the other arm with the same labor.
Generally chronic disease management is very labor-intensive and requires a lot of follow-up with patients, and that’s a big part of the reason why it’s been so challenging to move the needle on outcomes
HCI: With the formation of the company, was Stanford the first place that you tried to roll this out more broadly?
Vakili: Because Ashwin [Nayak M.D., M.S., co-founder] and I are on faculty at Stanford, it is one of the locations where we are deploying this system, but it’s important that you have different environments and different patient populations, so we have a handful of systems that are involved beyond Stanford
HCI: You have Allegheny Health System, UCSF and Cleveland Clinic. How did you get them involved?
Vakili: I think the fact that I’m a practicing primary care physician and being able to work with other clinicians and team members on this systemic challenge is key. We’ve been in stealth for years, and I think showing them this vision galvanized the group. We see the potential of this, and people are excited to do this in the right way. I think it just comes from the network of physicians that we have in the community.
The other thing, too, is that I was a VC before running UpDoc and that definitely helps. I’ve had many health system relationships beforehand. I’m also an M.D./M.B.A. Like a lot of my colleagues who went to school with me, there’s a particular breed of primary care physician M.D./M.B.A. who wants to try to fix the system by tackling these challenges. Those colleagues of mine who are like-minded end up at many of these different institutions
HCI: Is there some skepticisms at each of those health systems that has to be overcome before they approve and allow this to be rolled out? And some governance processes to work through?
Vakili: I wouldn’t call it skepticism. We’ve been proactive and very much believe you need to have adults in the room for a system like this to be done responsibly and safely. Because we’ve been so proactive in having those conversations up front with them, it’s helped those systems feel more comfortable. And the fact that physicians are involved with this at every layer of the organization, it’s helped them understand and think about and ask questions that they wouldn’t normally even think to ask.
I think there’s a lot of respect for the fact that we have done the hard work of going through the FDA process and we’re trying to approach this as thoughtfully as possible from that perspective. They also appreciate the fact that we still have physicians double-checking the system, because in real-world environments, there are always unknowns compared to study environments. We are working in collaboration with the systems to be able to think thoughtfully about the governance frameworks and oversight frameworks. We’re in no rush. We think with a long-term view, it’ll all come into place.
HCI: You mentioned working on this six years ago. That was before a lot of the large language models took off. Has that affected your platform?
Vakili:Six years ago is when we worked on this in the research realm, and the technologies have just further improved and developed. The tools to be able to have conversational interactions with patients were developing, and these frontier models get better and better and better. It very symbiotically just helps improve clinical AI systems to be able to operate more effectively and be able to have a more seamless experience with patients.
HCI: Is there integration work that has to happen with the health systems’ EHR platform? Does it have to be customized to their own specific clinical guidelines, which might be different at Stanford vs. Cleveland Clinic?
Vakili: Yes, EHR integration is mandatory for the system, because it’s very important for the clinical AI to be embedded in clinical workflows that the patient is going through and for the care team to be aware of what’s going on and how you are working collaboratively with the clinical AI tool. I would say 90% of the EHR integration is the same across systems. There’s about 10% that’s different for the practice environment, but that’s pretty typical for large health systems that have their own needs and workflows.
HCI: We’ve talked about this initially as a sweet spot for primary care, but what about nephrologists or other type of specialists who are helping patients with chronic conditions? Couldn’t it work just as well in those settings?
Vakili:Yes, it works in any chronic disease setting. We’ll definitely be talking about that in the coming months. I think primary care needs the most help in some ways. In my medical school class at Johns Hopkins, not a single person went into family medicine, not one. I think I’m one of two primary care physicians from my residency class. We’re all internal medicine. We could have all been primary care physicians, and I think I’m one of two in my graduating class. So we just need a lot of help. And I think if these tools are used thoughtfully, they could help primary care physicians spend more time with their patients.
