Dan Keller 0:02
Dan, welcome to this episode of Substantial Matters: Life and Science of Parkinson's. I'm your host, Dan Keller, at the Parkinson's Foundation. We want all people with Parkinson's and their families to get the care and support they need. Better care starts with better research and leads to better lives. Most people with Parkinson's disease will experience gastrointestinal problems at some point in their journey, making the gut microbiome a big focus of Parkinson's research over the past decade. Today I'm speaking with movement disorders neurologist Dr. Lisa Deuel of the University of Vermont Medical Center. She explains how the gut and brain are connected and discusses common gastrointestinal problems in Parkinson's, such as constipation and slowed stomach emptying, called gastroparesis. She also shares practical strategies to support gut health. So I first asked her to put in perspective the gut microbiome as it relates to Parkinson's. 

Lisa Deuel 1:20
The microbiome is a term we're hearing a lot more, and what that refers to specifically, with regards to the gut, is the microbes—so, any bacteria and viruses that essentially live, in this case, in the gastrointestinal tract. You can have a microbiome other places in your body as well, but when we think about the microbiome as it pertains to Parkinson's, we're often talking specifically about the gastrointestinal tract. And I think of it, you know, we're almost trained to think of bacteria and viruses as being bad and being associated with illness and infection, but this microbiome is actually part of what we call a healthy gut, and needs to be there in order for the gut to work properly. It's there from birth, and there's a lot of different things that can shape the microbiome. So, just even the way that you're delivered at birth can impact what your microbiome ends up looking like; your genetics can play a role in what your microbiome looks like, and as we'll talk about, we know that the microbiome tends to look different in people with certain diseases, including Parkinson's disease.

But the microbiome is really important because it helps our bodies to develop a healthy, working immune system. It helps with the way we digest food and absorb nutrients and vitamins from our food. It actually—some of the bacteria produce chemicals that actually help to communicate from the gut to the rest of the body, including the brain. So, it gives signals communicating what's going on in your gastrointestinal tract, and then, you know, how that impacts the rest of your body. So, it actually not only impacts just the gastrointestinal tract, as we would assume, but it really has connections with the entirety of the rest of the body.

Dan Keller 3:16
And a normal microbiome seems to keep things pretty well in check and functioning properly. People without it get into trouble.

Lisa Deuel 3:26
Exactly. We have data to suggest, again, not just in Parkinson's, but in other diseases as well, that there may be changes to the microbiome. Where it gets really tricky is trying to figure out, you know, are those changes happening at the beginning, and therefore causing the disease, or are these changes that we see as a result of the disease? And we don't necessarily know how these changes impact the start of a disease like Parkinson's disease, mostly because we have a lot of data to suggest that the changes going on from a molecular level in Parkinson's disease are happening long before a person actually develops the classic symptoms that we think of, like tremor and walking trouble and slowness of movement. So, it's hard to know exactly how these changes to the microbiome actually cause or affect the disease process.

We just know when we take a group of people with Parkinson's, for example, and we look at the pattern of the bacteria that we see in their gastrointestinal tract, that it looks different than people who don't have Parkinson's disease. Again, it's not just one specific bacteria that is or isn't there; it's more so the pattern, and that the pattern in Parkinson's disease tends to be more inflammatory. So, these types of bacteria are more likely to cause inflammation or irritation of the gastrointestinal tract, and so when we start to notice these patterns, and we see more susceptibility to inflammation, there are theories that this then makes the gut leakier. Perhaps that's an entryway for toxins, potentially toxins from the environment, or things that we ingest to enter the body.

Dan Keller 5:28
There seems to be at least two hypotheses: gut-first or brain-first. The things you see in the gut are a result of what was going on in the brain, or things that precede symptoms may start in the gut and move up to the brain. Where does that stand?

Lisa Deuel 5:48
I think that's exactly right, that there's these two ideas going on, and there's data out there that supports either of them. And it may be the case that we're lumping a lot of people into one category called Parkinson's disease, but maybe there are actually different processes, different disorders that end up with the same symptoms, and we're limited by our ability to diagnose people based on what we see clinically. We can't really categorize people by what's going on inside the body, and so I think that's where it stands, is there's limitations based on how we're able to obtain data and characterize people.

But there is a lot of interest in looking at the microbiome, looking at people who, let's say, have a family history of Parkinson's, so may be more likely to develop symptoms in the future, and really try to categorize these things as early as possible, so you might be able to tease out the difference: is it coming more from the gut or is it coming from the brain, or are there two separate, let's say, populations of people? So, maybe both theories are correct, and again, that's where there's a lot of interesting research going on to try to help answer these questions, and really find out, could any of this, whether it's the microbiome or other chemical changes in the body, could those be considered a biomarker that if we look at that early in people who don't have Parkinson's disease, it might predict whether or not they're going to develop symptoms?

Dan Keller 7:31
Can changes in the gut affect the symptoms or progression of the disease?

Lisa Deuel 7:37
Yeah, it seems like modifying the microbiome can affect Parkinson's symptoms. I think whether or not that's changing the disease process itself or changing how your body can, let's say, absorb and utilize medications is kind of a question up in the air. What I mean by that is the way that people have looked at this so far is to, let's say, in some current research, give people a probiotic. And the idea with a probiotic is that we're helping to build up more of the good, healthy, less inflammatory microbiome, and then these small studies, so far, have shown that that can be really effective in improving Parkinson's symptoms.

But I think a question that still remains is, are we actually changing what's going on with the Parkinson's as a whole, or are we changing the way that the gastrointestinal tract is able to kind of break down and absorb medications and thereby just making the medications more accessible, so somebody feels better? I will say, too, there's plenty of data out there suggesting that something like a probiotic can help with the gastrointestinal symptoms—so not the motor symptoms necessarily, but people dealing with constipation, which is a very common issue in Parkinson's, that that can actually respond to a probiotic as well. So, there's a lot of different ways that somebody with Parkinson's could benefit from, let's say, a probiotic to change the microbiome.

Dan Keller 9:14
I always think that if our microbiome changes, whether we've had a powerful antibiotic or some other condition causes it, maybe we should have banked stool samples when you were healthy, sort of like doing cord blood transplants.

Lisa Deuel 9:32
Sure, there's interesting research, and maybe a little challenging to talk about, but there's some interesting research starting with animal models, with mice models, where they can effectively transplant the microbiome from, let's say, a healthy mouse into a Parkinson's mouse model, and you can see a reduction in Parkinson's symptoms. And they do the opposite as well, where they take a mouse with a normal, healthy gut microbiome, and then they implant the microbiome from a Parkinson's mouse, and again, you can induce more Parkinson's symptoms in mice models. So the idea there, in practice, it's challenging, but this is actually a process that's done that's called fecal transplant. It is done more for people with severe diarrheal illnesses, where you're trying to clear out that bad gut microbiome for somebody that's having really severe gastrointestinal symptoms, but it is effective as a treatment.

There have been small studies that have shown if you do a microbiome transplant through a fecal transplant in humans, you could also improve symptoms as well. If you really get into the weeds of this, it's perhaps a little bit not palatable in a certain way, but they can transplant the healthy microbiome. So, some people would find that type of treatment to be a little bit questionable, but there have been small studies that suggest it works. The trick again with Parkinson's is, you know, this is not something that we're doing ahead of time, and most people who develop Parkinson's don't realize early on that that's something they're going to develop in the future. So, unless we could come up with some sort of marker that would indicate somebody's risk for developing Parkinson's, you would just have no idea that you needed to do this in the future.

Dan Keller 11:30
Gastrointestinal issues can be pretty common in people with Parkinson's: constipation, gastroparesis, which is slowed emptying of the stomach, and they can be frustrating to people, and also embarrassing to talk about, even with their physician. So, how common are these kinds of symptoms in people with Parkinson's?

Lisa Deuel 11:51
You're right, they're very common. Some studies show up to 80% of people with Parkinson's have constipation; that's what we tend to talk about the most. But you're right, there are other symptoms related to the gastrointestinal tract that can be present as well. It is really hard, right? These are things that we're not used to talking about frequently. I'll say in my office, that's something I talk about a lot. So, I've gotten a little bit more comfortable over time, and a lot of these, you know, what we refer to as non-motor symptoms—it's important for us, as the clinicians, to ask about them, because people don't necessarily realize that it comes along with Parkinson's. Again, we often think about the motor symptoms, and people aren't aware of just how common some of these non-motor symptoms, like constipation, can be.

So, my approach is just be open about it and ask questions to try to get to the bottom of what's going on. You know, at times it makes people feel a little uncomfortable, but ultimately, you know, we're all just trying to work together to help somebody feel better. And oftentimes, I'll say constipation is a great example of that, that sometimes these non-motor symptoms can be worse than some of the motor symptoms that we're used to dealing with. So, I've learned over the years that if I don't ask, if we don't have some way of capturing the information, either through our intake forms or by asking directly, that this is a symptom that somebody might not offer up because of embarrassment or just not realizing that it's related. But it's really important because it's not something that you're necessarily seeing a gastroenterologist for. Most people don't see a gastroenterologist routinely for mild constipation, so I find in my clinic it's one of the common things that we're managing and treating.

Dan Keller 13:50
Constipation would be considered a non-motor symptom, but in essence, doesn't it have to do with the motility of the gut, the muscles, the smooth muscle of the gut propelling things through it?

Lisa Deuel 14:03
Sure, that's a good point. The term motor symptom we generally refer to as more of the mobility, the tremor, things that to some degree are visible as affecting the muscles, but it's a good point. You have a whole separate nervous system within the gut. The reason I think it's worth differentiating, though, is often the treatments are very different. So, when we give medications to treat the common motor symptoms of Parkinson's, we're thinking about stiffness and slowness and tremor. We have a number of different medications we can use to help with those symptoms. Those medications that help with the typical Parkinson's symptoms often don't help these other non-motor symptoms, and in fact, sometimes the medicines we use to treat Parkinson's can actually make non-motor symptoms worse, like constipation. So, there's a number of Parkinson's medicines that we would give to, say, help with a tremor, and we actually have to counsel people to watch out for worsening of their gastrointestinal issues.

Dan Keller 15:12
What are some practical steps that people can take to improve their gut health?

Lisa Deuel 15:19
We have a lot of different things we recommend, and generally starting from more conservative to more advanced, depending on the situation. In general, though, I would tell everybody, I guess, with Parkinson's or without Parkinson's, is to stay hydrated, drink plenty of water. I tell my patients, obviously, depending on the situation, even up to one or two liters of water a day, and that can really help with a lot of things, but constipation, too. And then eating a diet that's high in fiber, so trying to get in as much fiber as you can, can help with some of these symptoms that we're talking about.

We talked about probiotics a little bit, and how that can be a positive impact on the microbiome, and introduce more of those good, less inflammatory bacteria. So that's an option, especially as I said, there are small studies that say it can be helpful for constipation as well, and the other things we always think about is just eating healthy in general. Being more active can also be helpful here, and vice versa. The better your gut health, the better you're feeling. The better you're absorbing your medications, the more easy it is to be active as well.

Dan Keller 16:41
Since we're talking about the gut microbiome, it seems there's a lot of commercial companies now offering home gut microbiome tests, but there seems to be no standardization, no regulation, not even proficiency testing for the different companies offering different collection methods, different processing methods. Some are looking for DNA to know what's in the gut, some are looking for RNA or even metabolites of gut microbes. Then you've got the issue of sending in the sample through the mail: time, temperature, or bacteria dying off or proliferating. So, is there any utility to these kinds of tests, are they worth doing, and then if so, what do you do with the results once you have them?

Lisa Deuel 17:25
That's exactly the big question, is what do you do with the results, and I think to take a step back, it just reiterates what we've been talking about, which is, you know, we know the microbiome and Parkinson's, and again in other neurologic and non-neurologic disorders too. We know the microbiome is different, but we don't know exactly how it's different or how it gets that way, and so I worry about right now being in the midst of this research and not having really good conclusions about what the data means, is that you spend a lot of money on some of these tests. And it is really interesting, the different things that they can test nowadays, but it's hard to really know what to do with that data, so I generally tell people right now to not spend the money and maybe invest more, like we talked about, in just overall better diet and gut health by way of diet and exercise and hydration, rather than spending a lot of money for something we don't know how to interpret.

Dan Keller 18:28
As research in this area marches on, what gives you the most hope about the future of gut health knowledge about it as it relates to Parkinson's?

Lisa Deuel 18:39
I think the reason that it interests me so much is because we're still trying to answer this question, of course, of what causes Parkinson's disease. Is it truly one disease, or is it multiple different causes of different diseases that ultimately look the same downstream? And we've heard before about this idea that genetics plays a role, but the environment plays a role. This is a way to kind of tie those things together in a sense, and I think that's why I find it fascinating is thinking about how the environment might impact your gut microbiome, and then have a downstream effect.

And then if we can come up with a pattern to use the microbiome as a biomarker for people being at risk for the disease, that's something you can look for; you can look for that pattern in people without Parkinson's to really quantify their risk of developing it. So, I think it's interesting to me to fold in that there's some evidence if you adjust the microbiome, whether it's through diet or these different experimental capacities that I talked about, you could make somebody feel better in the moment, but also if you look more into the microbiome, you might be able to predict who's going to develop Parkinson's down the line, so that you can intervene a lot earlier.

Dan Keller 20:07
Very good, I appreciate it. If you'd like to know more about the gut microbiome in relation to Parkinson's disease, you can search for "microbiome" on our website at parkinson.org. You'll find several blog posts, news articles, and a past podcast. As always, our PD information specialists can answer questions and provide information in English or Spanish about this topic or anything else having to do with Parkinson's. You can reach them at 1-800-4PD-INFO. If you'd like to leave feedback on this episode or to let us know what other topics you'd like us to cover, visit Parkinson.org/Feedback. Be sure to subscribe and rate us on Apple Podcasts or wherever you get your podcasts. At the Parkinson's Foundation, our mission is to help every person diagnosed with Parkinson's live the best possible life today. To that end, we'll be bringing you a new episode in this podcast series every month. Till then, for more information and resources, visit Parkinson.org or call our toll-free helpline at 1-800-4PD-INFO. That's 1-800-473-4636. Thank you for listening. 

Lisa Deuel, MD is a Neurologist and Movement Disorder specialist at the University of Vermont Health Network in Burlington, VT. She completed her Neurology residency at Albany Medical Center in Albany, NY followed by a two-year Movement Disorder fellowship at the University of Colorado in Aurora, CO. 

In her current role, she provides comprehensive, patient-centered care for individuals with PD and other movement disorders. She has particular interest in gastrointestinal symptoms associated with PD, especially the relationship between gut health, environmental exposures, and inflammation on symptoms and disease progression.

Dr. James Beck 00:00:00
Hello everyone, and welcome to the Parkinson's Foundation Expert Briefing. I'm Dr. James Beck, chief scientific officer of the Parkinson's Foundation, your host for today, and it's a pleasure to have you with us. In today's briefing, we'll explore how artificial intelligence, or AI (something maybe no one's ever really heard about), and these AI-enabled tools may influence communication, symptom tracking, care personalization, and how individuals and care partners can engage with these tools responsibly. But before we begin, I always like to get a sense of who's joining us today.

We're going to launch a poll. In this poll, if you're joining from Facebook Live, you can respond using the comment section as part of the process. So, in this poll, you'll see what we're really trying to do is get an understanding of your connection to Parkinson's disease. Are you a person with Parkinson's? Are you a spouse, care partner, family member? What's your relationship? This helps us really understand who's joining us today and can allow us to tailor our presentation on the fly.

I'm expecting a lot of people with Parkinson's joining, but sometimes we get surprises with a lot of scientists and researchers as part of the process. So, let's see what we get as part of that process. Hopefully people have gotten a chance to respond. You can do it on your phone or, if you're sitting in front of an old-fashioned computer like I am, use the mouse to click the right spot as part of the process. Let's give it another minute. Okay. Yeah. Great. A lot of people with Parkinson's joining us today, as well as some other health professionals and some scientists and researchers.

Great. Wonderful to see. Thank you for sharing your connection to Parkinson's with us. Before we begin to get into the meat of our discussion today, I always like to take a moment to introduce the Parkinson's Foundation. We are a nonprofit organization dedicated to improving the lives of those living with Parkinson's by enhancing care and advancing research.

Our efforts are deeply rooted in the collaboration with the Parkinson's community and ensuring that everything we do aligns with your needs and priorities. Today's program is an example of how we're working with you to meet those goals.

For those of my colleagues, we've got two slides presented right here, so we just need to eliminate one screen. There we go. One of our key initiatives is part of this, you know, relies upon feedback and engagement of the community, through investing in breakthroughs. PD GENEration is, I think, top of that list. Here with PD GENEration, we offer free genetic testing and counseling to people with Parkinson's.

By participating in research, you can learn about your genetic connection to Parkinson's disease and contribute to the research effort that drives new treatments and, hopefully, one day a cure. So we encourage everyone to share this opportunity with your community. We know a lot of people in our Parkinson's Foundation family have already participated, so thank you very much for that. I think the key thing here is that together, as always, we can make a difference.

Dr. James Beck 00:03:56
And so what I'd also like to do next, and I think it's very relevant for today's topic, is introduce you to Ask PAM. This is the Parkinson's Foundation's latest tool, the Parkinson's Assistance Messenger. PAM is an AI-powered chat tool from the Parkinson's Foundation that provides trusted, evidence-based answers about Parkinson's disease. You can ask PAM anytime, anywhere. The key thing about PAM is it's secure, it's confidential.

As always, it was created to give people with Parkinson's and their caregivers instant access to accurate Parkinson's Foundation-rooted information 24 hours a day, seven days a week. The information that comes from this chatbot is all derived from stuff that we have created and vetted. It doesn't go out beyond the Foundation's network.

And so you can really feel confident that the information you're getting is something that's accurate. Like all the information that we provide on our website, this is really educational information. It's never meant to replace medical advice from your healthcare provider. Use it in that way, of course.

One thing we always do with our Expert Briefings is record them. This is no different here. So if you are having to step away, or you come in midway, or you know people who want to watch this, there will be a recording of this. If you've signed up, you'll get a link to when that recording is ready, and then at that point, you can share that with others as part of the process.

Now I'd like to take an opportunity to introduce our speaker for today, Dr. Allan Wu. Dr. Wu is a professor of neurology at Northwestern Medicine's Parkinson's Disease and Movement Disorders Clinic. He's worked as a physician informaticist for over a decade on electronic health record projects with the specific goal of looking to improve patient engagement, reduce physician burnout, and enhance efficiency. He collaborates with the Parkinson's Foundation today and the electronic health record provider Epic on a national initiative to improve hospital safety for those living with Parkinson's.

This work has led to the integration of PD-specific safety features into Epic's EHR system, helping ensure that people with PD receive safe and consistent care wherever they're hospitalized. Boy, I'm a tongue twister today for me. Dr. Wu, thank you for joining us today, and I look forward to your talk.

Dr. Allan Wu 00:05:17
Thank you, Jim. I'm going to go ahead and share my screen and get going. That was a great intro, and a privilege to be here. All right, great.

Thank you. My topic is artificial intelligence and Parkinson's: understanding promise and pitfalls. I'm going to get right into it. There's a lot to cover. My disclosures: I actually don't have any financial disclosures with anything that I'm going to be talking about. I have volunteer disclosures that Jim touched upon. I volunteered at the American Academy of Neurology. I'm chair of the Practice Management and Technology Subcommittee and a member of the Medical Economics and Practice Committee, and I'm also a member of the Brain & Life editorial board, which is a free publication online for people with neurological conditions and their caregivers.

Also, I am a volunteer elected member of twelve other neurologists around the country on Epic, which is a major electronic health record vendor steering board, which helps put neurology-specific content into that electronic health record vendor.

We're going to go through three parts. The first part is a basic understanding of AI, defining what it is and how it works at a very high level. Again, people on this call probably have differing experiences with AI. So those of you who are experienced, you'll have to put up with this part. And for those of you who don't quite know and are curious about AI, maybe you'll find something here to level set. Then I'll go through some examples of AI, broadly speaking, in action, real-world examples that are relevant to our community, and then some basic tips on how to use AI, particularly the large language models, safely and responsibly.

But first, a little bit about me and why I'm here, perhaps. First of all, I do not consider myself an expert in AI. As will become plain, I use parts of AI and have some experience with using it, but what I really am, as in my introduction, I am a physician informaticist. I'm a clinical informaticist. I will define that more in the next slide. But I've been involved with this world since 2013, and that's a long time ago, and that predates the modern era of AI.

Dr. Allan Wu 00:07:43
There's a slide there, a picture from 2014, where there was a group of us mid-career physicians at UCLA that formed a group that advocated for health information technology during a transition to electronic health records, which were promised to solve all of our healthcare problems, and guess what? It did not. Therefore, there became a retained need for physician expertise in health information technology and making it work for us and not against us.

There became, in 2013, 2014, bottom left, the American Board of Preventive Medicine has now certified a physician specialty in clinical informatics. I am the program director of that at Northwestern now, and we train physicians to become stewards of electronic health records, health information technology, at the level of helping individuals. That is physicians, that is patients, that is caregivers, and that is our staff one-on-one with health information technology.

The things on the right I already mentioned in my disclosures, but I did want to give you this idea that I've been working with health information technology for a while. Clinical informatics is not just about the technology. It is about the people who are using it. It's about the process of making sure that it's integrated into workflow. These pictures and headlines are from my other talks on clinical informatics, and they are talking about how we need to make sure that the little picture on the left is showing the busy doctor on the left, typing busily on their computer, on their chair, back against the very happy kid on the exam table and their parents.

It was a commentary on how we have become, we are typing on computers more than talking to our patients. The promise was that electronic health records would improve our care and solve a lot of problems. It did not. AI is promising the same thing, and in my view, I think it's still about people, process, and technology. I do not think AI is going to solve all our problems, and it's still going to require people, advocates, physicians, and physician informaticists to extract the value out of them.

Dr. Allan Wu 00:10:20
This is a picture of what's called the fundamental theorem of clinical informatics. This is from 2013, and it was the premise that a human plus health information technology should be better than the human. That is a driving theme and should be the theme as we move into a new era of technology, including AI. And what defines better? This is the classic kind of triple aim of healthcare: improving the patient experience, lowering the cost of care, and improving the health of the population. There are some that will advocate for a fourth aim, which is reduced clinician burnout, because if we have physicians retiring in droves, that does not help our efforts either.

But this is background about where I am coming from into this world. With that, I'm going to go into the first part, which is the AI 101 understanding AI. There is hype about AI and a promise, and these are quotes of very, very optimistic, overly so, people.

Demis Hassabis there on the left is a computer scientist, but he won the Nobel Prize for using AI to show how proteins could fold. He's claiming that maybe we'll cure all disease with the help of AI in the next ten years. And on the right side is a prominent essay from futurist Dario Amodei, who said, I think AI could increase the rate of our discoveries by ten times. I think this is more in the mark of trying to enhance our abilities with AI technology, giving us all this progress.

Of course, you get these headlines. The first one, from a pilot at Sutter Health, saying, you know, with ambient AI, which is AI listening to the conversation and writing clinic notes for the doctor, doctors can give patients their full attention. There's reasonably good evidence that that is actually helping occur. And then, of course, HCA Healthcare System was prominent in trying to leverage AI to focus on improving safety in their hospitals.

Dr. Allan Wu 00:12:44
Those are the hype and the promise, but there are at the same time these concerns that the ambient AI is going to weaken your relationship with patients. That's from the AAMC, representing medical schools and colleges. AI chatbots provide poor answers to medical questions half the time. If a doctor was providing you poor answers half the time, I hope that doctor is not going to continue getting new patients.

And of course, Americans, and again, I understand there is a very broad audience here, international included. Some of the data I'm going to show in the polls tend to be based in the United States. There was a poll that basically showed that Americans are very concerned about AI, with one of the major concerns being the misleading video and audio deepfakes. You can see Sanjay Gupta of CNN over there basically calling out that there was an ad going around saying that he discovered how Alzheimer's could be reversed with a traditional simple Indian recipe, and he's saying, you know, that's not me, and this can only get more realistic over time. So there are concerns about this technology.

The reality is, just like electronic health records are in our medical clinics, it is going to be hard in most of the world to find clinical care without some electronic health records. It's harder to avoid AI in some form. Americans use AI, in this other survey broadly defining AI-enabled products, you see there, but basically 99% of us touch something: a website, a shopping site, our smartphone, our Amazon lists. These things are all included in that category. So it's here in various forms.

Generative AI, which is the new era of creating AI that is creating novel outputs, this is the chatbots and so on. Over half of people have used it for personal purposes, 28% at least weekly, and still about a third, 31%, never used it. The skepticism in this poll, which was March 2025, so between 2024 and 2025, it's already a year old, there was a sign of growing skepticism, with the feelings of being overwhelmed, skeptical, you can read it there, curious is a good one, cautious, concerned. And there was a bit of a shift downward in being impressed, hopeful, or excited.

Dr. Allan Wu 00:15:26
With that, I'm going to go into a little bit more about the 101 of artificial intelligence. There's a lot of definitions that get pretty technical, but basically I try to simplify it. It's basically computers doing things that, if a person did it, you would think they require intelligence. And I can hear you all thinking, yeah, yeah, there are things that we do, that humans do, that do not seem to require intelligence. But in the end, basically, computers are doing things, responding, writing back texts, talking back to you, and they are things that you would not have expected computers to be doing, say, just a decade ago.

So that's a very broad definition of artificial intelligence, and it comes in a spectrum. We're going to come back to this a little bit more, but basically, coming on the left, they are very, very simple rule-based types of assistance that computers can do. Rule-based is like following a flow chart. It's like an if-then statement. Flow charts have inputs, like if your cholesterol is high, get on an anti-cholesterol medication. You know, very deterministic. If the input is known, you have the predictable output. You can explain it. Any human can read it, understand it, and at most, flow charts are, what, one to five decisions deep.

Statistical models get into about ten to a hundred different parameters, right? That's like how many things do you need to put into your statistical model to predict how, you know, a table for life insurance, for example. What information do you need? It's not going to be ten things. It might be fifty things. Machine learning increases that again into the hundreds to hundreds of thousands of parameters.

Dr. Allan Wu 00:17:13
Machine learning is taking patterns of things and learning from those patterns. This is a workhorse that you'll learn about what AI is doing in the medical field, and it's like your smart watch detecting AFib, right? It's finding patterns, but hundreds, a hundred thousand parameters in that.

Deep learning is kind of like machine learning, but different specialists are in there. So it's like processing a Face ID, where one machine learning thing detects the eyes, another the mouth, the other the face, and so on and so forth. And so that gets into the millions of parameters. Then generative AI blows it all out of the water. It now is in the billions of parameters, with rumors of the newest models having trillions of parameters. It boggles the mind. So what that means is the outputs are no longer explainable.

There is no way, like we can understand a statistical model. If you are older and have more comorbidities, diseases, you might have a shorter lifespan than somebody who is younger and healthier. But in generative AI, there is no way to even predict what's going to come out of that thing. More bottom-up, emergent, you know, quote knowledge. You can't predict what's going to happen. It looks novel, and it's been trained on lots of things. But you can sort of see the development of this from one side to the next.

Then I want to talk about this concept. This is from the American Medical Association, and I do think this is a very, very good framing. It's called augmented intelligence, which is basically defined as artificial intelligence that focuses on the assistive role of AI. Its design and its use should enhance human intelligence rather than replacing it. This basically is a reframing of the 2013 fundamental rule, fundamental theorem of clinical informatics: let the health information technology enhance and assist us.

And I think this is some theme that you should hold on to as we explore AI and as you explore AI. Again, AI 101, we unfortunately have to go through more definitions. You might come across these. I don't think we need to dwell on this for the purpose of the talk. But you might come across these things in your reading, and in media and lay materials.

Specific AI is AI, any of the types that I showed, that is trained to do one specific task, like this AI tool transcribes a doctor visit into a clinic note. That's all it does. It doesn't do things like plan your vacation, right?

But a general AI could do things on novel problems that it is not trained on. So the example here is: what did my doctor mean by dyskinesia? You can say, explain it to me in Spanish, or explain it to me at a sixth-grade level, or give me dyskinesia in the context of all the movement disorders that could exist. It'll answer things beyond what it's been trained on.

Generative AI actually will combine that into more novel output, where it will summarize things and create a new kind of information with follow-up questions. And the agentic AI are basically AI agents that are empowered, meaning you gave them permission, if you are in control of that agent, to take actions. So you can give it permission to go to your computer, reorganize your files, clean up your inbox and your email.

Then the example there on the side is, you know, this is sort of a more future approaching, but it reads your visit summary. It will then reach out to your pharmacy, request a med refill, submit a prior auth for you, schedule a follow-up with you, and keep calling the clinic for you on your behalf over and over again until a cancellation opens up and you can get your appointment, that kind of thing.

So, just to put a pin in it, and we can go through this quickly, rule-based are things like spell check and appointment reminders. Is the word spelled right or not? Is the appointment today? And if it is, send a reminder, right? Very rule-based, easy to understand.

Probability models, those are your insurance risk calculators. Your weather forecasts are probability models. It changes as time goes on, likelihood. Appointment cancellation is an example of application in clinic operations.

Now machine learning is those Amazon, Netflix show recommendations, a very fancy text completion that happens automatically in some emails, and patterns in your wearables and your Oura ring or your smartwatch. Deep learning is more deep, multi-large data extraction tasks like dictation, speech recognition, voice-to-text, Face ID. Photo sorting by some of the photo apps will now sort by face. That would be deep learning.

And then generative AI, we will talk about ChatGPT, is the prototype that came out in 2022. Then there's these other ones, Claude, Copilot and chatbots. Now you've got PAM at the Parkinson's Foundation that is a generative AI product as well.

Dr. Allan Wu 00:22:35
AI can mean any of those things, and it can be more than one of those things. So, if you just take the example of the AI, ambient AI in the office visit, a microphone is listening to the physician, the patient, the family and the conversation. That's deep learning. Deep learning is looking for patterns in the audio waveform pattern of what it's hearing, accent or not, and it will process that and turn it into text that a computer could compute.

Then generative AI will take that information and combine it and summarize it into a clinical note. Then a physician is responsible for reviewing that note and approving that it is their responsibility, and that would be the augmented intelligence part of things. But, you know, we will say ambient AI and we mean all of this.

Okay. Large language models. So this is the most accessible part of the current era of generative AI. You can see, I tried to figure out a way to represent how popular it was. This is a Google Trends search that I did using the term ChatGPT, which basically doesn't exist until the end of 2022, and then it starts taking off. It's like one of the most popular. People signed up in droves very, very fast.

But it is very expert in language processing, right? It got trained on text information, and it is fluent and it is confident. It's prone to inaccuracies. I'll go into it a little bit more. It simulates. It is not truly conceptual understanding. It's basically learned from all the text that it could get trained on, which represents a lot of, I mean, humans are very prolific in explaining and writing down text information, like a university library across the world.

That's where it's basing its knowledge on, and so it produces really fluent language. It is a very good language processor. So conversational tools that answer questions, summarize, explain and draft, that is an early entry point for generative AI. But of course there's also other content generation, and you've seen all these things online in the news. It can create images. It can create videos. It can create slides, spreadsheets, documents. It'll create computer code. Not going to go into all of that, but all of these count as generative AI and has a lot of capability.

But I want to emphasize that while these are accessible, and I didn't land on an exact analogy, but we'll build one here during this slide here, we want to make sure the human is in the loop. Large language models are a tool, and I would treat it as a team member that enhances our capabilities. It should never, ever replace decision-making, human connection and judgment.

Dr. Allan Wu 00:25:45
Humans are always, always going to be needed. And the idea is not just to shrink your workforce, right? It is to design a better team. It is not necessarily for a smaller team. We can do more, hopefully, if we use it responsibly and remember the tool is a tool. It is a computer. It is not a person, and it doesn't have any responsibility or real knowledge.

It is really, really good at being a super librarian. It will organize and summarize large amounts of information. All the information in the Parkinson's Foundation websites, which is quite a bit, is now accessible to you through PAM. It's very good at broadening your searches, broadening your horizons, broadening your brainstorming, providing multiple perspectives and options. That is what it's good for. It's like the super Labrador retriever. It will retrieve anything you want, as long as you tell it to broaden the search. It will do that.

However, these are designed to be overly agreeable to a predictable fault. So this is like the super graduate student that is able to retrieve anything you want, but basically has no common sense, has no sense of responsibility and takes no responsibility for mistakes. It's overly agreeable. It will get anything you want. It has amazing capabilities as a super librarian, but it will have no real pushback on anything you ask. And so you have to guard against that.

And so this has also been prominent in the news, but this is the reason why we use the term hallucinations for generative AI, where it just makes up output. Bias is very hard to detect. It reflects gaps and assumptions baked into its training data. Certain genders, underrepresented topics or people, not just people, but topics, are a form of bias.

So, for example, if you have Parkinson's disease and you are asking questions about it, well, there is a good amount of information out there about Parkinson's disease. But if you have some rare genetic syndrome, and I'm sorry, I'm going to make this up. I don't know, ATP17 G2 genetic disease, and there's only five or 10 of them in the world, then there's not much training data there for that. If you ask it questions about that entity, it will give you just as much information as the Parkinson's is, because it made it up, because it is so agreeable. That is a bias that, you know, if you didn't know that it was that underrepresented, you may not be able to tell.

And then AI slop has come into the lexicon now. It's copious AI-generated content without any human effort, often with generic, limited meaning and quality. So my advocacy to you is: be aware of this. Please don't contribute to it.

Dr. Allan Wu 00:28:55
Last slide on the AI 101. The reality is between the hype and the disaster. I think that the tools are, I've shown you that they're available to you. You can use them. There is, on this YouGov survey on the bottom there in purple, 19% of people did not use tools at least weekly. In the top bar, 30% of citizens have a baseline positive, somewhat positive view of AI.

But of those who use tools at least weekly in some form, again, there's a chicken-and-egg issue here, but in general, I'd like to think positively on it that if you dip your toe in the water, if you're using this at least weekly, you start to understand that there are ways to use this that are positive. Hopefully, this talk will encourage that 19% to 30% of you who are thinking about using it will take advantage.

All right. Part two, we're going to shift into some real-world examples of AI-enabled tools for people with Parkinson's and their caregivers. We are going to talk about symptom tracking and monitoring. We're going to talk about wearables and passive data collection. We're going to talk about communication, education, research, clinical applications and then a quick wave at the future here.

Symptom tracking, one of the first things that comes to mind is, you know, when you interact as patients or caregivers with the clinical care team, that is a very small percentage of your overall life. You have symptoms that occur daily, and it would be wonderful if we could incorporate those symptoms, your tremor, your movements, your cognition, your thinking, your impact on function, what you're doing every day, into the care plan in some way, and get good, reliable medical advice on how treatments can best be evolved.

This has proven to be tricky, right, because there's not a standard way to do this. So AI has the promise of integrating that together and producing graphs, summaries and so on. But the platform really has not coalesced in any one single way. So what I'm going to do is just show you different examples, a list of examples of where this is being done. But I think that this has a ton of promise, and it already is here.

Dr. Allan Wu 00:31:34
Many of you have already been doing this. There are standard rating scales. There are patient diaries that make it a little easier to integrate, at the cost of, you know, it is hard to put the information in because you have to remember to do it. But these patient portals and clinics and health systems will send surveys out to some of you. Some clinics will do that. Some don't yet. There are other various third-party symptom trackers like the APDA symptom tracker.

And there's a bunch of FDA-approved devices now that, for the most part, are monitoring tremor. They're monitoring other things, and many of them will also collect your own symptoms. Now, once you symptom track, or you spend the time entering it, and of course people have been doing this for a long time even in Excel sheets or pieces of paper, right? People do that. It's putting this together in a form that the care team can use, that you can use, grab insights from and take action on. Those are hard, and that may be something where the large language models and the generative AI can help. And machine learning is good for finding patterns.

So let me go into a little more detail here. This is an example of one of the, and I'm not endorsing this one over any other one. I was looking for a good example of illustrating how symptom trackers work. This is a free app, StrivePD. Probably many of you are using it, and if you haven't, feel free to look at this or any of the other apps that I will mention. But it's an app that you can manually log your symptoms, program in med reminders. It might, I'm not even sure, but it might even integrate with, say, your smartphone's health app as well.

It will sit there and learn the patterns and start to even find ways to remind you to do certain things. Then it will integrate with your smartwatch. So Apple Watch is integrated into this product, and it will monitor your tremor or your dyskinesia, and it will detect that passively without you having to enter in data itself. But if you do enter data in, like your symptoms every day, your mood every day or your med reminders that you've taken your meds, that information is all in here together in this app, and machine learning will start to derive patterns and can create graphs, summaries over a week, over a month, what's on average, and those sorts of summaries can be helpful for your care team.

And then they have another tier, and I think you have to pay for it, where it advertises an AI companion or concierge-style experience with human support. So this is obviously a large language model, generative AI, that is added on top of machine learning insights, where it will summarize things in a more concise way that might be easier for you to understand, use and query for new insights. I love the fact they offer human support because, as we're going to find out later in the talk, these generative AI things, the deeper you go, the more challenging they are to make sure that they are being completely factual and helpful.

Dr. Allan Wu 00:34:47
This is an example taken from their website, which is symptom log data recorded from symptom logs, daily check-ins and notes. It says it's generated by AI, but it's literally, it looks like something that you could just give to your doctor instead of your doctor asking you: so, have you had balance problems? Have you fallen? Do you have off periods? Do you have dyskinesias? And getting responses, and they dutifully transcribe it into the note. This could just literally go in the note. As long as the clinician reviews it and it's short enough that a clinician can read it, it can be in the note.

And then we get into this era, which I favor, which is patients and physicians co-creating the clinic notes. But that's a different topic. This is an illustration of how AI can be helpful. Again, not meant for treatment. That's where the word wellness is from. This is not vetted, and so you use it for advice, but it is not medical advice.

Okay, wearables and passive data collection. So I learned a lot in researching for this talk. There is a lot, and there's more than just what's here on this slide. These things all monitor different parts of potential symptoms that people with Parkinson's would be affected by: gait and tremor and cognition and sleep and speech. I guess the smart clothing and shoes are kind of on the gait spectrum, with smart socks and so on. Some of these will do some treatment or symptom management things, like the vibration for your gait or neuromodulation in the case of the Cala device there.

The asterisks are for devices that are FDA approved. But the FDA approval is usually very limited. It's for, you know, detection of tremor is the most common one, even though it collects data that can be used for other purposes.

So, all of these things, the message here is these are all machine learning based. Machine learning is the workhorse because it is capable of discerning patterns out of large amounts of data, and it's still enough on the left side of the spectrum where it has somewhat of a predictable output.

Dr. Allan Wu 00:37:15
If somebody is wearing a smartwatch and it detects AFib, and if you have the same EKG pattern, then AFib is reliably output. That reliability does not exist for large language models and generative AI, which is why generative AI is not what these devices can market as at this moment. None of them really do, but they all say AI-powered. You just have to understand, these are machine learning, which is not bad. Machine learning is incredibly powerful, but it is learning patterns, and it's not the same thing as what you're doing with the generative AIs in creating images and documents.

Now, on the other hand, for communication education, that is where large language models start to come into play, because we can do things like, can you explain the doctor's note to me? Can you explain the symptoms, diagnoses, test results and medications? Again, it does not replace medical expertise and context, because it will explain these things in a very generic way, and it may or may not apply exactly to your situation, though I will show you some ways to mitigate that bias and try to make these things a little more useful to you and your care teams.

I think interpreting medical bills and EOBs, whatever that is, right, explanation of benefits, can be helpful. Again, be careful what you put in. We'll talk about that later. Be careful what you put into these products. They're all learning from your data and information. Drafting letters, questions and summaries for you, your care team or your billing team, if your medical bill looks a little off, who knew. Pre-visit, prioritize your agenda with questions, translate your notes after the visit and create an action plan. I'm going to give you more on large language models in a bit.

There are also patient portals with chatbots, smart speaker assistants and website chatbots. These things are evolving, and they are helpful, and you can use them. You just want to be alert, even if it is provided by a clinic or a medical center that is very famous, it will still make mistakes, and you want to be on your guard for that.

Clinical applications: for this group, I want to just note that AI is starting to be used in clinics for clinical purposes, and you might come across these in your experiences. Ambient AI is probably the one that has the most penetrance into clinical settings because it has been shown to increase face-to-face time, save physicians time in writing notes. You should know that, and you should feel free to ask any clinics that do this, where is my data going? How is it handled? And in most cases, there will be a reasonable explanation of that. Some will de-identify, say, a part of the data collected for future training, and you should realize that you can opt out of these things if you wish.

Dr. Allan Wu 00:40:32
You might be sending patient messages through a portal, and the clinic may be using AI to triage those appropriately. People will write messages for all sorts of things because that is the nature of the portal. I need to change my appointment. I need this pre-authorization. My billing is wrong. I need to ask the doctor this clinical question. I'm in the emergency room, and I have these questions. So the urgency varies, the purpose varies, and the type of person who can respond to you effectively and efficiently is going to be different.

And we have seen clinics where the staff are not empowered to do anything, and everything goes to the clinician, burning out your clinician. Others are, trying to get through the staff is like getting through Fort Knox to get to the doctor's response. So trying to use AI to help triage that is an active area of development. Drafting initial replies sometimes occurs as well. That's gotten some more mixed response. But for the most part, patients will find that, gee, doctor, your responses seem unusually polite and professional. Are you okay? Well, it's because AI actually drafted in its wonderful fluency. They're still responsible for their content fully.

AI-assisted letters and notes. The next one, AI-assisted review of patient-reported data and remote monitoring data, is the machine learning and the large language model interpretation of the devices that I talked about earlier. That's an evolving space, but very active because of so many devices that are sort of in research. And I should say a lot of those things I put on that slide earlier are research and not actually generally available. I wanted to illustrate that this is an evolving space of monitoring symptoms.

And then an amazing opportunity is AI-assisted clinical trials matching. It really has not matured yet. NIH TrialGPT is a really interesting research project at NIH that showed a demonstration project for this. But, you know, Parkinson's Foundation, Michael J. Fox Foundation have, you know, Fox Trial Finder, trial matching services. They're not AI-assisted at this point. This is still an evolving opportunity.

Then, I'm not a researcher, so I'm not going to talk a lot about this, but suffice it to say there is a lot of AI working to assist in interpretation of these big data sets, including the PD GENEration, lots of opportunity there to identify patterns in that data set. So we appreciate everyone, patients, caregivers, controls, who are participating in these studies.

Dr. Allan Wu 00:43:35
Then future, I mentioned it once before, and I'll mention it once more here. Agentic AI is coming down the pike, really has taken off in 2026, particularly for writing computer code, but these are AI agents that act in the real world.

Some examples, and this has not really taken off yet as relevant directly to people with Parkinson's, but these are the ideas: it's continuous passive monitoring through voice, gait and typing patterns and acting on it and alerting you for things. So, for example, proactive outreach when the wearable data signals a change; adaptive medication reminders based on your on-off patterns, and depending on whether you missed a dose or not, or you woke up late, those sorts of things happen. You stayed out late. What should I do? Symptom triage, helping you decide what needs a call, what can wait, and then all the stuff I talked about earlier in the example where AI can reach out and help you schedule automatically, help you follow up on referrals to make sure they don't fall behind and that they get followed up with.

And this all leads to sort of our holy grail of precision medicine, where it's not just about genetics. It's about making sure that your treatment is tailored to you. If we do have the genetic information, which would be incredible, because everyone has a different type of Parkinson's most likely, that we tailor that treatment. We want to personalize patterns for your medication and deep brain stimulation programming, for those who have heard about that. It's a bit of trial and error, and wouldn't it be great if you could just make sure all the DBS programming is kept in mind as the next programming is done?

Especially accounting for the fact that Parkinson's, unfortunately, still is progressing. What worked before may not work as well now. How do you optimize that? Personalized exercise coaching and the learning healthcare system, where your data improves care for everyone, and your data can also improve care for you, feeding information back into the point of care.

Anyway, this is a lot of future stuff. I want to turn ourselves back to sort of practical reality now. Oh, yep.

Dr. Allan Wu 00:46:04
So, now we're on practical tips. That was sort of an overview of things that are in existence now at kind of a high level. I want to address the current generative AI and the chatbots, right? And so we would love your feedback on this kind of thing. But large language models are accessible entry points to this world of AI. There's a lot of free opportunity to use this. You can hardly avoid it if you use Google search, but you have to understand that prompting and interacting with these things are not search engines. You can, and people do, use them as search engines, but you get pretty generic responses.

And my point has been that we want to leverage this tool to be a useful team partner during your Parkinson's care journey. So, I keep using different, shifting metaphors because I didn't settle on one, but maybe you're getting the idea: very smart, capable partner with vast knowledge, overly eager to please and no sense of responsibility or common sense. So how do you corral this so that you're not led down a biased path and you're getting information that is reasonably trustworthy?

I have six tips and tricks. Again, those of you who are experienced at these may have come across all these already and more. But these are six that I chose that I thought could be helpful for a broad audience. The first one is, instead of using it as a chat, as a search engine, you want to use a standard prompt structure. I do try to think through this each time. Sometimes I get lazy and I don't, but it is helpful to remember it knows nothing, has no context.

So you have to tell it who to be, tell it what you want done, give it some context, give it the background it needs, and then tell it how to deliver the answer. You might say, help me prepare for my doctor appointment, maybe a few more than that. And then you can apply it this way and say you are telling the AI, you are a knowledgeable Parkinson's patient advocate. That immediately constrains the training set into the things that Parkinson's patient advocates kind of would say. Then help me prepare the three most important questions for my neurology appointment Tuesday. So that's what you want.

Then you give it some context, and there's some sort of medical generic context here. I even give it a website for the Parkinson's Foundation guide to preparing for a medical appointment, and I tell it how I want the information to be given back to me. This already helps constrain and give you more specific information right off the bat. It's sort of a good fundamental start.

Dr. Allan Wu 00:48:57
Now, I did say provide a context, but sometimes you don't know what the context is, especially if you are not an expert. For example, if I was using AI to help me plan a house remodel, well, I don't know the first thing about house remodeling details, right? I will have a general contractor come in and I will talk to them. But in the meantime, you can just say things like, you know, I want to plan a remodel and this is the kind of things I want to do. But before you answer, ask me any questions you need to help give a useful and comprehensive answer. That allows the AI to get the context it needs to be very specific.

So, I need to talk to my neurologist about whether it's time to consider DBS. You could ask that. However, if you just throw in this phrase, before you help me, before you respond, ask me questions you need to know to give me useful guidance, it might say things like, how long have you had PD? What medications are you on? How well are they working? What are your biggest symptoms? What are your concerns? What does your care partner think? Maybe I would have asked that question with context, and I would have thought of maybe the first two or three of those questions, but maybe I wouldn't have thought about giving number four and five there as context. But once you give it those things, they should give you a more specific answer.

Guide the output exactly if you can. This is where you can shape the output the way you want it to be. The basic idea is things like, give me a bulleted list, give it in lay language, respond at a sixth-grade level, respond in Spanish. Be very concise. I love the "be very concise." These things have a tendency to go on and on way more than I want. And so I have found "be very concise" is always helpful. But then you can go to the next level, and you can give it an example.

I will use Word templates or Excel templates and say, I want your response to fill in the blank cells in this Excel template. That worked well for a house remodel. A Word template, or you can just type it in if you want, but a Word template is like, I want a business plan. Here is the business plan format I want.

And you can also tell it how not to respond. That is also very helpful. For example, in full disclosure, I don't use AI for these slides. I will use AI to give a slide to AI and say, please give me a strategic opinion about what I put here. But these things have a tendency to want to just create a slide for me. I'm like, stop doing that, right? You know, respond with a text description and I will create all my own slides. Then you can coach it for better response. So you can go through this a couple times. You can tell the AI, review your response, how would you improve it?

Dr. Allan Wu 00:52:02
And it will sometimes give you a little more information. So this is a case where I said, hey, you help me prepare for the next visit, and I basically, please give it to me based on the Parkinson's Foundation guide for preparing for a medical appointment. In most of these tools, you can give it a website and it will read the website, and it will help you follow that output.

Okay. This is super powerful, and this is very helpful in sort of helping discern and protect against bias, right? I would routinely strive to ask for broad perspectives. I don't think it's really helpful, necessarily, to just have a conversation with the AI chatbots because, just like the overeager graduate student with no context, they will respond with one answer, which is whatever they think is most prominent in the training dataset. Then you respond to that, and they respond with one answer back, and you have no idea if they're leading you down the wrong path or not.

So, ask for more than one response. Ask for responses from more than one perspective. You can use this to have it role play for other things as well. Examples: respond as a professional physician, professional coach, professional strategy business administrator. Respond as a strategic consultant with deep knowledge about Parkinson's disease advocacy. The things I do a lot of is I say, give me pros, cons and alternatives.

I will sometimes say, ask for five different responses and options, and it will give me five responses, and that gives me an idea of the range of things that I'm dealing with. That's why I say it's very good for brainstorming. So you can say, first level, respond as my movement disorder specialist. What would concern you the most? And so on. You can simulate that.

This is super helpful for difficult conversations as well. I give an example down here on the bottom right. You can just say, I was asked to discuss the possibility of a feeding tube. I am not sure how to begin. If I bring this topic up to others, respond in turn with the perspectives and questions my neurologist, social worker, psychologist, pastor, family member and my patient would have. It will give you lots of perspectives, and it gives you an idea of what you're dealing with before you sort of zoom in on being more specific.

Dr. Allan Wu 00:54:37
And this is also a really, really helpful trick to keep in your tool bag, and you get caught by this. This is called show your work or show me step by step. This is an example that happened to me earlier this month. I did an analysis of some data in a spreadsheet, and I took a picture of the spreadsheet after I did my analysis. I just wanted to know if I was missing anything or if anything else the AI would pick out of the data set. So I said, tell me what you can conclude from these numbers. It said, that seemed a little more optimistic, didn't seem right.

So I went ahead and said, "Hey, I was thinking that it was a different answer. My interpretation was different. Please re-review your work." And it came back and said, "Oh, I'm sorry," and it gave me another conclusion that was aligned with what my interpretation was. But you could tell it wasn't quite right. You know how you can tell when the student is not quite right? They're like, "Please give us a book review of To Kill a Mockingbird," and they say, "There was a mockingbird and it was being hunted." You know, you didn't really read it, did you? Right. So I said, "Walk through your thinking step by step. Tell me the first set of numbers you worked with."

And then it said, I can't really read your spreadsheet. It came through blurry. It made up the whole thing, confidently, apologetically and wrongly. This is called chain-of-thought walk-through, thing step by step with actual details. Do this if you're doing anything that is going to be concerned about accuracy.

Memory, I think this is my last point. Memory, when we have conversations, is very important, and it's limited to each conversation you're having with the chatbot. So you want to save your work. You have to make sure it doesn't remember context. Every new chat you have, it remembers random things actually, but you want it focused on the context that you care about. So create a permanent context for each topic. These are called projects in Claude, custom GPTs in ChatGPT, agents in Copilot, terrible word. But I have a context for how to set up my internet system at home. I have a context for my financial planning. I have a context for my home remodel, right? So I don't have to repeat myself each time, and I can start a new chat and ask a different question.

Dr. Allan Wu 00:57:06
I don't have to keep repeating the context. And then be careful: if you're having a long conversation, like I rarely have a conversation with the chatbot that is longer than 15 minutes, 30 minutes at the most. Long conversations get distracted, lose focus, and inconsistently use parts of the previous conversation. It can be subtle and you may not even tell, but it starts to obsess over certain things. You can just stop the conversation and ask it to create a prompt.

I will usually read the prompt and change it a bit, but it's not bad. And you create a prompt and you can use it to start a brand-new conversation. It's a good trick.

I wanted to mention NotebookLM. It's a personal file cabinet. I have nothing to do with Google, unless I use Google. I actually don't use this tool a lot, but it promises that it only answers from things that you've given it. So it tends to have a lower hallucination risk and it can tell you if it does not find something. You can upload things like your visit summaries and test results, and then you can interact with it and ask it questions about your own personal items. I know we're running along, so I have basically two more slides to finish up here.

This is a potentially useful tool to look into, but I'm not endorsing it myself. I think it has promise. Two slides on responsible AI use. One is privacy in your data. Free tools are going to use your conversations for model training, and anonymous health data can still be re-identified. So please be careful. Don't put in things that are sensitive. Consider starting out, dipping your toe in the water, by paraphrasing things instead of uploading your whole clinical notes.

I think it's okay to take screenshots of parts of your clinical notes that hide identifying information. That's fine. You'll get used to what you can and can't do, but just be aware of what's there. When AI is used in medical settings, you can ask about opting out of the tool and what happens to your data. HIPAA still governs your health information. Anyone that is using AI for your healthcare is responsible for safeguarding that information. And you are responsible for safeguarding your information by all those things that you give permissions to on your website and on your phone as well.

Lastly, I mentioned the bias. AI tools can underperform, so please use some of the tips and tricks maybe to mitigate that. Legal accountability, it's still an evolving landscape. If someone has a harm from a diagnostic error, who is responsible? Right now, it's mostly the physician. So the physicians are responsible for what comes out of these AI tools if they're using it for advice. FDA oversight is very limited. It's only for medical devices, and almost everything I've talked about is not under FDA oversight because it's for non-diagnostic purposes, or they're consumer health wellness things, and they're largely unregulated. So, stay informed.

AI is a tool. It's evolving fast, and there are opportunities and some risks. Use it thoughtfully. Don't create AI slop. Know the limitations. Keep yourself in the loop. Start small and be curious. Ask questions of your care team. Contribute to the Parkinson community on how to use these things. I'm sorry I ran a little long, but I wanted to leave time for questions, so that's what I had prepared for you. I threw in some websites for things that you can use to get started. The Coursera course gets referenced a lot. I verified the hyperlinks, but not the individual courses. I think that's it.

Okay. Thank you.

Dr. James Beck 01:01:13
Thanks, Dr. Wu. That was really quite a tour de force. I mean, lots of information to cover as part of this process, for sure.

I just want to say that before we start taking the questions, just know that our team is organizing them as they come in and that the topics that we really have time for are the ones that are going to be, you know, that Dr. Wu addressed in his presentation. Nevertheless, we'll try to get to as many as we can, and know that my colleagues here are trying to do that. We also have, as will be mentioned later on as we close, our Helpline, which is a fantastic resource as well as part of that process.

Alan, it looks like you use AI a lot, but seems like there's a lot of effort to do that. They hallucinate, they get obsessed on small things, they can lose focus. I didn't realize that. I mean, I thought this was a computer and that was the advantage of computers, that they don't have these issues. So as you said, they show great promise, but they have these great pitfalls. How are you using them in your day-to-day life? You just mentioned you don't use them to prepare slides, but you use them to kind of make certain your slide makes sense, that type of thing. Is it something that you see professionals use more often? What about patients coming in as part of the process? Do they come in with what ChatGPT does?

I took your example and put it into ChatGPT for your knowledgeable Parkinson's patient advocate, to ask these three questions, and it came back with some ones that looked pretty decent as part of the process. So I'm just curious. There's a lot to unpack there, so let me give you a chance to respond.

And you're on mute.

Dr. Allan Wu 01:03:04
Yeah. I do use it every day. The way I use it, first of all, I do not think AI saves me any time. Now, I think that's very interesting because I have talked to a lot of fellow physicians that are using ambient AI, and it saves them tons of time. I actually use ambient AI as well in the clinic, and it does not save me time, because I actually review it very carefully. I'm a neurologist, and some of us neurologists are OCD and we want it to say what we really mean. But what it does save me is cognitive effort, right?

It is far, far easier to edit something that you care about than to write it from scratch. I think ambient AI doesn't save me time, but it saves me cognitive effort and leaves my brain alive to do other things. The other way I use, and I mostly use the large language models. I was on ChatGPT, and now I'm on Claude.

I interact with it, but I use these tricks that I showed you. I care about them because I shared them with you, and I really use all of them. I really do use context. I really use it to brainstorm. I really use it a lot to say, give me different perspectives, give me 10 different options, what are we missing? If I'm writing an email to, say, my chair that is kind of sensitive about various things, I will use AI to say, give me five different ways I would say this.

And then I will review those because it has to be my voice. And I will tell you, even though I use it and it remembers me, and you can give an example and say, this is an example of my email, use this voice, it does not do it perfectly. So I always end up writing my own anyway, but I will feed it back and say, here's what I wrote. What do you think? Give me three pros, cons, and alternatives, because I'm guarding against going down a rabbit hole.

So it's good for that, and that's how I use it. I don't use it for images and things like that, and they're never satisfactory to me either.

Dr. James Beck 01:05:25
Yeah. It seems like, as you described here in your presentation, it's really about how, if you're using these large language models, it's really about how you ask the question and making certain you're very specific about how you want that response and what they need to think about. You just can't go in and say, tell me about Parkinson's disease. You need to say, I have Parkinson's disease and I know something about it, so you can get to a different level as part of the process. It sounds like.

Dr. Allan Wu 01:05:57
Yeah, no, that's exactly right. I mean, with the modern tools, you can do things like, I have Parkinson's disease, I want to promote it, I want to be an advocate, create for me a website about Parkinson's disease, and it will do that. It will spend lots of energy doing that, but you weren't specific enough. It just spits out a generic website, right? It will answer like the no-context, no-responsibility thing that it is.

Dr. James Beck 01:06:25
Yeah, that's a good point. Hallucination is this thing to be on the watch for because it makes up answers. It seems to me that you're talking about it wants to please you and provide a response. It never seems to say, I don't know, which would be great. I would trust people who say, I don't know, because there's questions people just don't know. So how are you able, again from a perspective of a person living with Parkinson's who's trying to really get information, how do they know when it's hallucinating? How do they know when it's starting to make up answers when you don't have necessarily the context to come in with and say, oh, that seems bogus?

Dr. Allan Wu 01:07:14
Yeah, yeah. One trick I didn't say is you should always verify the source independently. It's sort of like getting your spam call and you happen to answer it because you were expecting a package from Amazon, and they got lucky and say they're trying to sell you something. So the advice is hang up, right? And then verify by logging in independently to Amazon or your bank and verifying it yourself. So what you do is you use these tricks I told you, which is walk me through this step by step, give me a range of options, give me the sources.

And then you don't click on the links in the chatbot, just like you would never do that in an unsolicited text message. You would independently open a browser. This is what I do, and then you go to that website and you read the website. Maybe I'm cautious, but then you go to that website and you read the website.

When I said at the end, Coursera course, I verified that link. It gave me a bunch of links because I said I don't take lay people AI courses. I wanted to give you guys something. So I asked it to brainstorm broadly. I asked it for like 20 different options. I picked out ones, and then I independently verified them. I will tell you what I independently verified was not what it gave me. I modified it and I changed it, so on the slide I was comfortable saying it because I'm responsible. I'm a human. I'm responsible for what I put on the slide. That thing is not. So verify independently would be the other thing.

Dr. James Beck 01:08:50
Yeah, for sure. Okay, some more questions that have come in as part of the process. A lot of people again are thinking about the availability of these tools out there, Copilot, ChatGPT, Claude. How do we know which one's more accurate than one? I mean, is it just catch as catch can, or are there ones that are really like, ah, this is the one to go for versus others?

Dr. Allan Wu 01:09:15
Yeah, I don't think I'm the best person to answer that. I think these are all evolving tools with varying levels of accuracy. None of them are perfect, for sure. Claude had some marketing things that occurred that increased its reputation. I think there are, and I don't know what the resources are, benchmarking tools for accuracy that are out there, benchmarking these things. I did not look at them because I equally trust and do not trust these things.

There is one called OpenEvidence that markets itself as being a little more accurate for physicians that want to research literature, and that's the reputation it has, but I can't verify that. And then if there's benchmarking tools, are they verified for, say, people with Parkinson's in the Parkinson's domain? That we don't know. One thing we could do as a community is say, here's the benchmarking toolkit examples that we could use to vet whether these tools work or not, right?

Then depending on the answers they give, followed by a group of experts, we could say this one works better for Parkinson's domain questions. But we don't have that right now.

Dr. James Beck 01:10:41
Yeah, for sure. I don't mean to put in a plug for our own AI chatbot, but I'm going to, and that's our Ask PAM. I think one of the goals here, and just put it in context if you will, I don't know if you've had an opportunity to see our chatbot, is that we worked very hard to try to minimize a lot of these concerns as part of the process. We've used what's often termed the walled garden. We give the AI language model free rein to go wherever we have already created materials, our website, all our stuff. It can't go beyond. This way, we've ensured it's a safe area, all the material has been vetted.

Often, if there's anything that's involved some kind of consideration about research or medical, someone has outside-reviewed it so we can ensure the quality of that. How does that compare to some of these other tools, which may see a broader realm? There's no walled garden, they get free roam of everywhere. What are the implications of that versus our PAM tool?

Dr. Allan Wu 01:11:51
Yeah. Well, I think the Parkinson's Foundation is not alone in creating these walled-garden things, right? There are a lot of specific AI applications of chatbots that are doing exactly that for different entities, different organizations. The American Academy of Neurology, for example, released a walled-garden version for us to navigate a very large annual meeting recently, and I can say that it did not work great. It had some limitations. That's why I mentioned NotebookLM at the end there, because that is what it promises, that it will wall it off for you.

You put information in NotebookLM and, yes, Google is probably going to learn from it, but it promises not to go outside the walled garden of your stuff. So if you put in, I don't know, all of your car maintenance visits, if you go to different car maintenance, and then you can query, when did I change my spark plugs? It will tell you, right? You can do the same thing for health.

But yeah, I think the other thing to look out for on PAM and these specific AIs is if you have an excessively long conversation, will it stay on track? That I don't know. I think most of these walled-garden things are discrete enough of a set of data that it can't get too distracted, but we don't know.

Dr. James Beck 01:13:19
So thinking about that, you had mentioned this idea of context, like a question that came in asking if contexts are something that you have to, like, are those part of the free version, or is that part of the paid version for some of these AI models? Do you happen to know? I imagine you're a big user, so you might pay for it, but I don't know if you know that.

Dr. Allan Wu 01:13:40
Yeah, I pay for one service at a time. But I think the free versions have these things because they do want them useful. If you use ChatGPT, look for , I forget what I called it. I had it on a slide earlier. Yes. And Claude has something called Projects. But yeah, context is, yeah, without context, it will give you generic stuff.

Dr. James Beck 01:14:09
Right, but my question, sorry, I maybe didn't ask it well enough, is if you have these contexts, is that not the same thing as a long conversation about the same topic?

Dr. Allan Wu 01:14:20
Oh, yeah. When I say long conversation, that means back and forth, say 50 times. You're having an active conversation with the chatbot. A context is where you can upload one large prompt and say, that's the context. You are an expert in the Parkinson's domain. You have access to everything in the Parkinson's Foundation, and you will answer in this. My goal is to construct a research grant to the Parkinson's Foundation for this purpose.

Then every time you start a new chat in that project, it knows that context as a prompt. Then you can have a conversation. Now, if your conversation goes 50, 100 levels deep and I've given it 20 different versions of a paragraph to react to, it's going to start getting confused. So I should start over and say, okay, I'm back at working on the method section of this paper. Here's, can you react to this? It will forget about the first version I did, but it will still have the context. You can have as many chats as you want within a context. You just don't have to type it all again.

Dr. James Beck 01:15:36
Got it. Yeah, I appreciate that. For people with Parkinson's, the less typing the better as part of the process. I couldn't imagine.

Dr. Allan Wu 01:15:46
They also take voice now. So most of the versions now will allow you, if you click the microphone, to just talk to it and it will transcribe the words using natural language processing. Then you can also set it so when it responds, it'll speak back to you.

Dr. James Beck 01:16:02
That's fantastic. That's going to be really helpful.

Dr. Allan Wu 01:16:05
Yeah, that actually is available in most of the models that I am aware of.

Dr. James Beck 01:16:09
All right. Well, at that point, I have to think we have to close. Thank you very much, Dr. Wu, for your time. It's been fantastic. Really appreciate that. A lot of information to share today. I'm sure we'll work on trying to get this turned into a fact sheet. Maybe we'll get AI to generate one for us from here. We'll probably actually get a real person and you as part of the process, but this is really useful information for our community. So thank you very much as part of that.

I know we had a lot of questions come in during our Q&A. Unfortunately, we could not get to them all. If your question wasn't answered, feel free to call our Helpline at 1-800-4PD-INFO, or you can email us at Helpline@Parkinson.org.

This finishes our spring series of expert briefings for the year. We'll come back again after a brief hiatus this summer and come back about disease modification and stem cells: where are we now?

There's lots of exciting information you've probably heard about in trials that are going on. Hopefully we'll be able to give you some feedback on an update of where the state of the field is as part of the process. Then, I think everyone's favorite coming up is the gut-brain connection in October. Then we'll close the year with environment and Parkinson's. This is an area that the Foundation is getting into, the idea of what could be environmental triggers for Parkinson's and how that might actually lead to ways to prevent Parkinson's for future generations, which would be really fantastic. You can always learn more and register at Parkinson.org/ExpertBriefings.

Our PD Health @ Home series is ongoing, offering a range of virtual programs, Mindfulness Mondays, Fitness Fridays, and you can learn more and register for PD Health @ Home by visiting the webpage Parkinson.org/PDHealth. Last, I would like to thank the Light of Day Foundation, who has always been a tremendous supporter of the Parkinson's Foundation and whose generosity here has made this programming possible. So thank you, Light of Day Foundation.

We're here for you. Again, you can reach out through our website, our Helpline and our toll-free number as part of the process. Until we meet again this fall, hope everyone has a good summer. Take care, and we'll see you on the next Expert Briefing.