AI-Based Analysis for Parkinsonism
11:48 Share

I'm Yulin Xun, Associate Editor of JAMA and JAMA+AI, and you are listening to JAMA+AI Conversations.

My guest today is Dr. David Valancourt, Orchard Professor and Chair of the University of Florida. His research program uses advanced neuroimaging techniques to study the functional and structural changes in the brain of humans and animal models that span Parkinson's disease, dementia, tremor, ataxia, and dystonia. Welcome, David. Hi, thank you for having me.

Can you tell me a little bit of history about your article, Automated Imaging Differentiation of Parkinsonism in JAMA Neurology and the reasons why you decided to develop this new AI tool?

Yeah, so in Parkinson's, you have an umbrella term called Parkinsonism, which encompasses Parkinson's disease and other forms of Parkinsonism as well that mimic the symptoms of Parkinson's disease, and that's where the misdiagnosis can occur. So in the first five years of a person's trajectory from a diagnosis, it can be anywhere from, say, 25 to almost up to 50% of the patients have the wrong diagnosis.

Once they establish the diagnosis, the physician establishes it over a long period of time, it does become much higher accuracy. But early on, within the first five years, the diagnostic accuracy is less certain. So we started on this trajectory because we're trying to

help physicians with tools that might be able to help them make these diagnostic calls earlier and more accurately because a lot of patients are being treated with treatments, which may be the wrong treatment. It's also mucking up clinical trials currently and in the past where patients are included in these large-scale clinical trials that probably don't have the type of

Parkinson's disease that these trials are looking for. And so if you look at drug development and you look at clinical care, we're trying to sort of improve those spaces because we feel as if a lot can be done to kind of make sure the cohorts and make sure the patients are correctly diagnosed. And your AI model looked at specific parts of the brain to learn patterns related to each disease. I found that really fascinating. Can you give a lay description of how the AI model did that?

Yeah, so we were able to distinguish these forms of Parkinsonism reasonably accurately across a really messy data set. So part of the problem in MRI is the manufacturers don't necessarily talk to each other or work with each other. So they are competing with each other in the marketplace. So their software doesn't always align. So having an analysis software that works across these scanner manufacturers is not trivial. Doing it on one scanner is a lot easier. But when you start combining it across data sets, that's the non-trivial part.

And so we're basically using an AI algorithm that weights all the reasons that are in our model. It's kind of like a fingerprint, you know, so you're basically trying to figure out where degeneration is happening in the brain. And the hypothesis going into the study is that certain types of Parkinsonism, so we're basically studying Parkinson's disease, malaria.

multiple system atrophy, and progressive supernuclear palsy. Those are different pathologies. They have different symptoms, but they can sort of mimic each other early on. The hypothesis going in was that these pathologies would have different signatures in the brain in terms of where the brain is degenerating. And so we're using an imaging technique that is picking up on those degenerative patterns. And then we're basically assigning them a diagnosis based upon those degenerative patterns.

And then you also compared the AI predictions with brain autopsies in some cases to see if it got the diagnosis right. Is that correct? Correct. So we had two different ways we compared the AI algorithm. The first was clinical diagnosis. And the second, as you mentioned, was autopsy confirmed cases. So we established the algorithm on clinically diagnosed cases.

And in the study that we published in JAMA Neurology, it was a perspective study. And that was the key part about that study is perspective, and it's across 21 sites. And we had all three major manufacturers for MRI scanners in that study as well. So we were able to test the AI algorithm across sites, but also across scanners.

And so the way that we defined the clinical diagnosis was we had three independent and blinded physicians look at the patient video, clinical characteristics, neurological exam, a standard MRI they had access to, and they could make their own diagnostic call independent of each other. And then when they agreed with each other, that's when we called that patient as ground truth, which would be the clinical diagnosis. Now, we also had a separate cohort of pathologically confirmed cases that we then compared the algorithm on.

So in that case, we had two different sources of ground truth to see if the algorithm was working in each of those instances. And if it worked in one but not in the other, then that would kind of cause me personally to kind of lose confidence. But if it's working in both, then that kind of gives us a lot more confidence that it could be a very good algorithm.

And the AI tool was really accurate. It was over 96% accurate in telling Parkinson's apart from MSA and PSP. Why do you think that's the case? Why do you think these algorithms are so good, and in particular, your algorithm AI tool?

I think because there are parts of the brain, and if you look in the literature, you'll see this, that are changing. So a pathologist, if you look at, for instance, PSP, and you look at the pathology studies that have been published, there are really consistent areas of pathology in these patients. So for instance, the superior cerebellar peduncle is a key part output pathway of the cerebellum. That is changing in PSP. Subthalamic nucleus is also degenerating in PSP.

Regions of the basal ganglia and also midbrain are also changing in patients with PSP.

Whereas in patients with multiple system atrophy, you'll get some basal ganglia changes, but you also get the middle cerebellar peduncle is another part of the cerebellum, which brings in information into the cerebellum from the brainstem. That area is degenerating in MSA. So, you know, it's not just one region. We're basically coalescing a lot of regions. If you look in the past at some MRI studies, they've really focused on one sign on the MRI scan or one region or maybe two regions, but we're taking a collective region.

And so we're basically processing something that the human eye really can't see. So that's why I kind of think it is like a crowdsourcing model for how the brain is telling us what these patients look like. And we're basically assaying many parts of the brain at once to see how the fingerprint or the signature is actually identifying these patients. So that's why we think it was more effective than maybe past studies have been.

And do you think this AI tool can be used in hospitals to help people get the right diagnoses faster? How do you think it can be integrated into care? Yeah, so I have a conflict of interest. So I own a company that has the license on this patent and also is applying for approvals from regulatory bodies in the United States. And so the vision of that company is to place it within the hands of the clinical care spectrum and within that workflow. So, you know, neurologist orders an MRI,

that typically goes to a radiologist or a neuroradiologist and they're going to do a standard MR read. Well, just basically using the scan, which is a diffusion MRI scan, it would then basically go through an automated image processing. There's a lot of different vendors out there that are looking at AI in the AI space and radiology. So you could think of this like a plugin within that workflow that provides a readout or a report back

to the neurologist or to the neuroradiologist. So the goal is to basically put this into the workflow so it's readily accessible and available. The image processing time takes about two hours currently. So that's certainly within the expectation of how a patient with Parkinsonism is going to be diagnosed. You don't need it faster than that for any sort of reason right now. So we don't think that's a rate-limiting time period in that case.

And do you think, you know, with these AI tools getting better and faster, do you ever see these becoming at-home tools for people to use to help manage their symptoms or self-diagnose, not just in the hospital setting?

I think AI, broadly, I could see patients having access to AI tools that might help them. I think in our specific example, ours is best in the hands of a physician that is trained to interpret it. I would be hesitant to put this in the hands of individuals, but I do think having it in the hands of physicians and then the physician making the call on their own expert opinion is the way to proceed with our tool. But I can see AI tools in general might have some place in the home in the future. And

And so what's the future of this AI tool? How do you see it going further? Yeah, so this specific tool in Parkinson's, we see helping in clinical care for patients. You know, there's 90,000 newly diagnosed patients every year with Parkinson's disease. And, you know, if you think 25 to 50% of those patients are misdiagnosed, I can certainly see that it could be quite useful.

And also we see it being useful in clinical trials. So you can imagine a drug gets close to being successful. And if you were to remove 10% of those patients who had the wrong diagnosis, possibly that drug is actually more likely to be successful. So we see it as being used in those two avenues.

And you mentioned that you looked at different types of machines for MRIs. Can you tell me how that imaging could be improved to have better data fit into the AI model? Is there anything you would like to see in those MRIs?

I don't know that we would change the type of MRIs that are being performed. I feel like there's always a push in the MRI field because time is money, is to always squash the amount of time it takes to acquire the image. So I feel like there can be a danger in that to

sacrifice quality of the image itself. So I do think pushing the envelope too much can possibly hurt quality of images. And certainly our processing pipeline depends on certain types of imaging. So we would not want it to move so far beyond what our software is actually able to read, if that makes sense. So we would hope that the imaging technology is still available from the scanner manufacturers that fit within the types of analyses that we're currently performing.

And are there any barriers that you see in the future that you would like removed in order for this type of AI tool to be adopted in the clinical setting? We're currently seeking regulatory approval. So that is a barrier currently. And then I do think there's a trust factor in healthcare settings and hospitals about sending data outside of their own data networks. You know, so a lot of these

Certainly our cloud-based infrastructure is currently sitting in Amazon, for example, and Amazon Web Services. So having a hospital system be comfortable with sending data in and out of that sort of a setting is going to be necessary for our case. We could adapt it to be in a different setting, but we would certainly prefer not to. So I think having this comfort level of data moving in and out of hospital settings for AI-type tools, I think is going to be a really important step going forward. So IT and hospital settings is really quite critical.

Well, thank you so much, David, for this interview. Thank you. I am Yulin Xuan, Associate Editor at JAMA and JAMA Plus AI. And I've been speaking with Dr. David Ballincourt about automated imaging differentiation or Parkinsonism. You can find a link to the article in this episode's description. And for more content like this, please visit our new JAMA Plus AI channel at jamaai.org.

To follow this and other JAMA Network podcasts, please visit us online at jamanetworkaudio.com or search for JAMA Network wherever you get your podcasts. This episode was produced by Daniel Moreau at the JAMA Network. Thanks for listening. This content is protected by copyright by the American Medical Association with all rights reserved, including those for text and data mining, AI training, and similar technologies.