AI and Wearable Devices: Potential and Pitfalls | Congress 2026

Many of us know Dr. Hausdorff for the incredible work he's done with respect to gait, freezing of gait, and fall risk in Parkinson's disease, and much of this research over the years has really been assisted by technology. And as such, he was tasked with the lecture entitled "AI for Digital Biomarkers: Potential and Pitfalls" as part of the plenary session that we [00:01:00] attended yesterday on AI in movement disorders.

He joins us today to discuss this topic. Thank you so much for joining us today.

Dr. Jeffrey Hausdorff: My pleasure. 

Dr. Mitra Afshari: Dr. Hausdorff, your lecture was really spectacular, and I very much appreciated that you approached this pretty expansive, large topic from a lens of freezing of gait and gait and balance work that you've done in your lab.

We're all very much aware of the exponential growth of digital health technologies and AI over the last even two to three years, and you even presented a slide that by the end of, I believe, this year, there'll be more than 2,000 publications related to this topic in movement disorders alone.

Dr. Jeffrey Hausdorff: Exactly. 

Dr. Mitra Afshari: Let's start by discussing the benefits of using digital markers to study chronic diseases like Parkinson's disease.

I believe our audience would love to know your perspective on the overall positive [00:02:00] aspects of using digital health technologies and AI to study changes in movement disorders.

Dr. Jeffrey Hausdorff: Sure. That's a great question. Do we have a few hours to discuss that? No, I think there's a number of different applications, and the bottom line is it hopefully improves the clinical care and improves the patient outcomes. How it can get there is in a number of different ways.

One is just provide more objective measures, more sensitive measures, so if a drug is started in a given patient, rather than waiting two years or a year to see its effect, it could potentially see a more subtle effect earlier on using objective measures. So that's one area in which digital health technologies can aid the entire process.

And I think digital health technologies in general can be augmented by AI.

 Digital health technologies are kind of sensor-based ways of measuring patient performance in various environments, and AI surrounds that to improve our ability to extract meaningful [00:03:00] information from it. So one area is, just let's say, improving the clinical exam in the typical physician's office.

Another application is telemedicine. We know that many patients either live remotely, or they're wheelchair-bound, or they're unable to get to the office for whatever reason. And using digital health technologies and AI can provide the physician abilities to objectively measure what's going on in the patient's life remotely.

So that's another really exciting area, and it basically extends the clinical office from the medical center to the home setting. A number of different technologies can be used. For example, lots of information can be extracted using a routine smartphone, whether it's using it as an input device for finger tapping and measuring the rhythm or the inconsistency with the rhythm.

There are a number of apps that have been developed to measure various parkinsonian symptoms. In addition, there's a really exciting line of work [00:04:00] using video, which can be captured from a smartphone, and basically the video can be turned into kind of a sophisticated gait platform.

By using simple captured video, we can run an AI model on it and then basically capture the movement of the entire body, all of the segments, and now we have exact movements. We know how they're changing over time, and we can quantify that, characterize that, and so it gives us a lot more information about what the person is doing at any given time, how they respond to medications, whether they're ON or OFF.

 So it's really an exciting venue.

Dr. Mitra Afshari: Definitely. Those are really great examples. I think we can all agree on those points of the benefits of digital health technologies.

Now let's turn to the pitfalls, which was a large component of your lecture as well. You identified several pitfalls in digital health technologies and AI, some more obvious than others. First, as you said, the I in AI can vary quite widely. Could you [00:05:00] comment on how you've come to appreciate that in your own line of research?

Dr. Jeffrey Hausdorff: Sure. Like you mentioned, there's a range. AI is a very broad term, and some machine models or AI applications are relatively simple. Others are quite complex, have multiple parts and multiple segments, and take a little bit of information and try and make a huge story out of that.

So the inference can be quite large at times. We've been very interested in measuring ON/OFF and motor fluctuations from a lower back sensor. We've done a lot of work trying to extract as much information as we can from physical activity, mobility from a lower back sensor that's worn continuously.

We know that it can help us aid and better understand what the person is doing at a given time, how they respond, whether they're active, whether they're couch potatoes, whether they're exercising. So it gives us a lot of information. But we haven't been so successful in really quantifying motor response fluctuations, which is a really important [00:06:00] aspect of the whole care of people with Parkinson's disease.

There was an exciting presentation this morning about a new drug that appears to reduce the OFF time, and they're still using the Hauser diary, which it's unfortunately kind of the state of the art. There's a fair amount of attempts using AI and combining that with digital health technologies to augment or replace the Hauser diary, which relies on self-report.

For example, there's some work using sensors placed on the wrist and combined with other sensors there. Looking, for example, at the amplitude of tremor and how that changes over time, presumably in people who respond to dopa, their tremor is reduced in the ON state. That makes sense intuitively, and there's some good evidence supporting that.

And there's more challenging work trying to decipher a similar story from a lower back sensor, where gait is certainly, in general, responsive to dopa, but it's not as one-to-one. So, their AI is a little more challenging and more [00:07:00] complex, but there is some work to support it. And I think we're likely to get there in the future.

Depending on how you look at it, there's some evidence that we're there already. But I would say more robust evidence will come down in the future. We in the field are missing really large, well-labeled, well-documented data sets that can be used to train the different models.

'Cause we know patients with Parkinson's are so heterogeneous. One person is tremor-dominant, one person is PIGD-dominant. They're quite diverse, and even the response to dopa will be diverse. Changes with disease progression will be diverse. So if we want a robust model that can capture all that, we need really large datasets, well-labeled, as kind of gold standard references.

And then once the model, if you put in huge amounts of data, the model can explore that, investigate it, and understand and decipher. The more that we put in a well-documented model, the output will be more robust and more reliable.

Dr. Mitra Afshari: Absolutely. The second pitfall I believe you identified [00:08:00] in your lecture, Dr. Hausdorff, was the question of generalizability. You provided an example of how your group has seen, for instance, the precision of freezing of gait detection by some of the digital technologies that you're using change depending on just the environment that the patient is in, that might not just be reflective of what's happening to the patient, but just the environment itself is affecting the data output.

For instance, the home versus the community setting. Can you speak a little bit more on that?

Dr. Jeffrey Hausdorff: Sure. Classically, researchers and clinicians have used freezing-of-gait-provoking protocols to identify and observe freezing. In the past before that, we relied on self-reports, and then typically what would happen the spouse and their partner would come into the clinic and one person would say, "No, I don't have freezing. It never occurs." The other would say, "Sure, you have it." And then, in front of the clinician, there would be no evidence. In part, that's [00:09:00] why it was named by Stanley Fahn as "the mysterious phenomena," because it came and go. No one knew. There was no rhyme and reason. To deal with that, there were these freezing-of-gait-provoking tests that were developed.

A large part, what we seem to know now, is that walking in a circle in 360 degrees back and forth, sometimes with a dual task, is one of the best and most sensitive ways of provoking freezing. It's pretty good. 

Dr. Mitra Afshari: I know that test well.

Dr. Jeffrey Hausdorff: So if we want to know if a person likely has or does not have freezing, that's a great way to do it.

Pretty short and can get a pretty solid answer. But it doesn't necessarily reflect the burden of freezing in the home setting. Most people don't walk 360 degrees left to right while dual tasking. And if they did, probably people with Parkinson's would find a way of compensating and doing their behavior in a little different way.

It turns out that it looks like the triggers of freezing and even the phenotype of how it behaves in the home setting and in these kind of artificial [00:10:00] FOG-provoking tests are quite different. And that means basically the model that we're training in one situation is seeing things very differently in the other situation.

In other words, we have two different datasets, in essence. One in the clinic, and one in the home. And to be able to capture both, we need data for both. At the moment, we have a fair amount of data on FOG-provoking protocols, and we have much less data in the home setting for technological reasons, for time-consuming reasons.

Since currently the gold standard of identifying whether the person had freezing is for an expert to review a video frame by frame, it's super time-consuming. There are issues of privacy. Most people don't want to have cameras in their home setting, in their kitchen. So, we're still waiting for these larger datasets that are annotated in the home setting.

And then once that does occur, I think it'll occur over time. There are ways of blurring faces and dealing with privacy issues. Then, the model [00:11:00] will be able to train and understand both what happens in the clinical setting but also in the home setting, which is more ecological and more reflecting of what's going on in daily life.

So that's a really important addition that we need to strive for.

Dr. Mitra Afshari: Absolutely. You describing it makes freezing sound even more mysterious.

Dr. Jeffrey Hausdorff: Yeah, it's quite a complex conundrum there.

Dr. Mitra Afshari: Absolutely. Dr. Hausdorff, one of the many hats you wear is an advisor to the MDS Technology Study Group. You briefly presented a recent investigation that was done by that group surveying movement disorder providers across the globe on their perceived benefits of digital health technologies and AI.

And in brief, the results of that survey, which are yet to be published, were that while many are able to see how beneficial the information provided can be for clinical practice, rarely are they, at least at this point in time, pulling those into actual clinical [00:12:00] practice. In other words, as you said, enthusiasm is high, but routine use is actually quite low.

What is your take on that identified gap, and what do you think can be done to close that gap? What do you think are the primary drivers? Obviously your answer may change in five years, but at least at this point in time, how can we get closer to closing that gap?

Dr. Jeffrey Hausdorff: Yeah, that's a great question. I think there's a number of issues that are contributing to this, and part of it also probably varies by country or medical practice and insurance. Whether we like it or not money plays a big part in the use of these applications. In order for these wearable devices, digital health technologies, and AI to become more commonplace, it needs to help the patient and their long-term quality of care but also needs to be super easy for the physician and the clinician.

' Because we know movement disorder specialists, like most doctors, are super busy, and every second [00:13:00] is super precious. If they need to learn a new technology, and if it's burdensome, or if they get a 20-page summary that it's hard to understand what X and Y means, then it's not going to be implemented.

So it needs to be super simple and meaningful and actionable. And so right now we have lots of data. Some of it is actionable, but it just hasn't really been packaged in a way that a clinician could look at it in three seconds and understand, "Okay, I need to increase the dopa. I need to decrease the dopa. There's too much dyskinesia."

That kind of needs to be happening. In addition, we need to find people who are going to pay for this technology. Technology is not free. While the sensors themselves might be relatively inexpensive or even cheap, while an app might be inexpensive and just be part of the phone, it still takes time.

In the US, for example, it needs to be charged to some insurance item. So that all is a process. In [00:14:00] addition to the technology and the AI and the machine learning, we need to change the ecosystem both for the clinician, the payers, and then I think things will change.

Dr. Mitra Afshari: Yeah, absolutely. I think we need to also give the clinician some protected time to train on these devices rather than expecting them to do it immediately upon entering the clinic. I think we'll close the podcast on that.

Dr. Hausdorff, thank you so much for taking the time to conceptualize your lived experience with these digital health technologies, both for our audience here at the Congress and then today for our podcast audience. Thank you. 

Dr. Jeffrey Hausdorff: My pleasure. Thank you. [00:15:00]