Podcast Transcript: Viz.AI: Improving Access to Stroke Care using AI

[Dr. Bryan Hartley]:
Awesome. So you said you came up with the idea for Viz at Stanford. What was kind of the genesis of this? What's the clinical need and what was the reason behind it? Was there a specific patient or something that prompted you to want to start this?

[Dr. Chris Mansi]:
Absolutely. So back then, this is in 2015, I'm in the Biodesign course with an endovascular neurosurgeon, Jeremiah Johnson, and the MR CLEAN data came out showing that thrombectomy was effective and not just effective, it had very low numbers needed to treat for a devastating disease that I knew well, and my grandmother had, had a large vessel occlusion stroke. And all of a sudden we had the therapy to treat a patient that was proven and not just in one trial, but in numerous trials in 2015.

But I realized that actually, if you looked at the statistics, very few patients who had an LVO, a large vessel occlusion, were being treated. It was around the 2%, 3%, and if a patient showed up to Stanford, it wasn't that they weren't being treated, that they usually were, is the fact that often if you show up to a hospital without an interventional specialist, without a stroke neurologist, it wasn't being picked up, it wasn't being referred.

Maybe it was, but the workflow wasn't appropriate or the ambulance would take too long to show up. There were a huge variety of reasons when you went through the Biodesign process, why these patients were A, not being treated as often as they should, and B, why in the STRATIS registry, which is one the largest sources of data we had at the time, we saw that the median time to treatment was between three and five hours. Whereas in the best centers you could get patients, if they showed up to the hub, you could get a patient from door to going in 30 minutes. And I spoke, you could get door-in, door-out times of an hour so why wasn't that happening consistently?

And I just thought back to some experiences I had in practice. I mean, I did the Ted talk about one young lady who had a subdural hemorrhage and who died despite the operation going well, and in the mortality morbidity meeting, realizing the reason why she died, wasn't anything we did in the operating room, it was the four to five hours it took and the delay to her to get her there. And you've been through the Biodesign program, it really encourages you to stop thinking about what you might do as the clinician and focus on the whole clinical need, because if it's the transporter that's the problem, if it's the transfer center that's the problem, that's where you need to focus.

So we realized that this new technology, deep learning, which had just come out of the labs of Geoff Hinton in Canada and was really all the rage in Stanford at the time, which was able to pick out patterns in data so it could tell you the difference between a cat and a dog, for example. We realized that potentially we could use that to pick out patterns in CT scans or in other medical data, and use that as a tool. At the time, there were a bunch of other AI imaging companies who were really focused on improving things for the radiologists, like improving maybe the measurement of something or the detection of something, but ultimately the radiologist does a good job of that for the most part. The issue is connecting the dots between the ED, the radiology, the neurologist and the interventionalist. With all of the amazing healthcare providers that go in between; the stroke coordinator, the stroke nurse, the ambulance driver, the EMT, and there's many, many people, it's incredibly complex.

But what if, within a minute of the scan, an algorithm would read that scan and alert the right neurologist or neurointerventionalist so they could go, "Yeah, these images show a patient who I could potentially treat, let's transfer them over." All of a sudden you would turn healthcare from a push model, with numerous steps and handovers and three to five hours of delays, to one where decisions could be made within minutes, where you'd have the team that really understands the disease and the time sensitive nature of the disease, driving the workflow, you could potentially shorten time to treatment significantly. That's ultimately what we have proven out.

[Dr. Bryan Hartley]:
So what I'm hearing, that's great, what I'm hearing is, it's really a workflow problem. From your end, it was a workflow problem and maybe the timing was right to use a technology like AI to help solve this problem. But again, it was probably, the workflow was the hard part for you guys; how do you figure out how to connect these dots and in a quick way? And then the catalyst would be once the scan is done, Viz would automatically read the CT scan right when the scan is done and then it starts this whole kind of domino effect afterwards?

[Dr. Chris Mansi]:
Correct, yeah. So it's reading any of the scans coming off the system. If it's a ischemic stroke, it's going to be a CT angiogram and if it sees a large vessel occlusion, a perfusion mismatch, it is going to send an alert to the relevant stroke team and associated vascular neurologist and the neuro interventionist, as well as the local emergency room physician and radiologist, and immediately they're connected, it's via mobile app. They click on the alert, they view the images in a matter of seconds and they can immediately communicate. So typically, we will see scan to decision times of around three to five minutes where you've not just got decisions of, "Hey, I can treat this patient," but you've got communication between that spoke in the hub where they're saying, "Okay, we're going to transfer this patient over to you name the comprehensive stroke center.

[Dr. Bryan Hartley]:
Yeah, that's great. I had a chance to download the training app that you guys offer and look at the images and from a radiologist perspective, the images were great, very high quality and it's a very intuitive user interface. I was curious how much time you guys put into your image quality and user interface experience?

[Dr. Chris Mansi]:
A lot, a lot. I think we realized early on that an algorithm on its own was not going to be the product and really, this would only work if the radiologist, the neurointerventionalist, the neurologist could roll over in bed and make a determination if this was a patient they needed to act on or not. So unlike a lot of the other AI companies, we kind of issued technological determinism, and instead leant into the complexity of how healthcare is delivered at 02:00 in the morning, as well as 02:00 in the afternoon at a major downtown center in Miami versus a rural spoke, and realized that actually the user interface was going to be the key.

In healthcare, and we learned this in Biodesign, the only way that you can change healthcare and workflows is if you make it easier for that clinician who's responsible for that patient and wants to do the best for their patient, you can only change their behavior if it's easier to use your product than the alternative. And the alternative [crosstalk 00:27:47]-

[Dr. Bryan Hartley]:
Also, lowering the activation energy, basically, that it takes because they have to make so many decisions on a minute by minute basis, that it has to lower that threshold for them.

[Dr. Chris Mansi]:
Yeah. I love the way you put that, exactly. Lower the activation energy, absolutely.

[Dr. Bryan Hartley]:
Okay. And so, as you say, the algorithm was not really... that's not the focus, you needed it to be a beautiful, almost like a work of art, which is what it looks like when you log in. It looks clean, cleaner than most PAC systems which was very refreshing. I am curious, how did you realize that AI was the way to go? I mean, I know it was coming out at the time and the timing did work. I mean, timing is so important for these things-

[Dr. Chris Mansi]:
Yes.

[Dr. Bryan Hartley]:
... but it also takes a lot to be able to say... first off, did you have any experience with AI? I know you said your mom was in computer science, but if you did not have any experience, that kind of takes a little courage to dive into something and say, "Hey, this is kind of the tool I'm going to use to solve this clinical need."

[Dr. Chris Mansi]:
Yeah, so I think we were looking at the system and we realized that you needed to focus on the workflow and the user experience, but you also needed a trigger. You need to try and replicate having a neuroradiologist in every CT scanner so that the neuroradiologist could trigger that workflow, because for every 1,000 scans that Viz processes, there's maybe 20 or so that actually will have a large vessel occlusion. So there's this 890, 980, sorry, but it's not [crosstalk 00:29:34], but we realized that the power of deep learning was to automate that focus. The algorithm doesn't need to, and wasn't designed to get to the level of a neuroradiologist. The point is these things, once you start looking at CTA, is a lot of relatively easy to pick up. But if you don't look at the scans, if you're an emergency room physician, they're hard to pick up, particularly the more distal occlusions.

So we needed something that could make care consistent, and we saw that in our FDA study. We were able, even in some of the best stroke centers around the country where we did the study, we were able to not only reduce the time from scan to alert of the specialist by around an hour, but most importantly, we reduced the variation from well over 120 minutes down to seven minutes, the variation of when this doctor was alerted.

[Dr. Bryan Hartley]:
Wow.

[Dr. Chris Mansi]:
To me, that's the major issue in healthcare that technology can solve. When things align, that patient is treated with the standard that we want them to be treated, but on average, that's not the case. There's so much variation in healthcare and AI solves the standard deviation problem.