MDJ Research Article of the Year: Cognitive Phenotyping Via Digital Analysis of Spoken Word Properties

This study explores a fascinating idea. When we ask a person with Parkinson's disease to perform a simple verbal fluency task, perhaps it is not only how many words they produce that matter, but also which words they choose and how they navigate through language, through the language they're producing.

To discuss this, I am joined today by Professor Adolfo García, from the San Andrés University, Santiago de Chile University, and the Global Brain Health Institute, who [00:01:00] is the senior author of the paper. Adolfo, first of all, congratulations on this recognition, and welcome to the podcast.

Dr. Adolfo Garcia: Thank you very much, Michele. It's a pleasure. And yes, this is a paper that we are very, very fond of, led by a brilliant student from our team, Franco Ferrante, and with marvelous input from the clinical side from María Fernanda López, from Hospital Posadas in Argentina, and a wonderful Latin American team.

So proud and thankful to be here representing the team.

Dr. Michele Matarazzo: Let's start with, what is the idea behind the study? The verbal fluency is something we very often use in our clinical practice and in cognitive assessments in general. Traditionally, what we do is basically we count the number of correct words that a patient can produce.

What made you think that there might be more information hidden in the actual properties of these words?

Dr. Adolfo Garcia: Well, because language is not solely about quantity. What you [00:02:00] say, the words that you choose, how you sequence them, all these things, are pregnant with information about how you navigate through language, which words prove more retrievable to you, and ultimately what your overall cognitive status might be.

This is a notion that we've been pursuing for a number of years now, and we first tried it and validated it in other disorders, including Alzheimer's disease, mild cognitive impairment. And then we had access to this wonderfully phenotyped Parkinson's disease cohort. Given the growing and by now well-established awareness that Parkinson's is far from only a motor disorder, we reasoned that there was potentially very valuable information in these fine-grained word-by-word characterizations of the output that patients produce to tell us a little bit about their cognitive skills, their cognitive profile, and a number of questions therefrom.[00:03:00] 

Dr. Michele Matarazzo: I like this step from quantity to quality of language. And when we talk about quality, there are some features that, to be honest, to myself and to, I believe, to other clinicians might not be immediately intuitive, which are some things like semantic variability, granularity, concreteness.

Could you explain these concepts in simple terms?

Dr. Adolfo Garcia: Yeah, absolutely. So when you think about a word, any word, let's say "dog," that word is a constellation of features. That word has many measurable aspects. Some features have to do with what we call sublexical features.

So words can be shorter or longer, right? So dog is only three phonemes long. It's much shorter than giraffe, for instance. Words also have a certain frequency of use, right? There are words that are very recurring in daily life. Table, for instance, right? And there are other words that [00:04:00] you don't typically encounter that much in your dealings with people in everyday life, right? Agnosia, for instance.

And then words vary in their conceptual precision. That's what we call granularity. So if I were to show you the picture of a banana, and I asked you to tell me what is this? You could very well say, "That's a thing," and you would be right. Or you could say, "That's food," and you would be right.

Or you could say, "That's a fruit," and you would be right. Or you could say, "That's a banana," and you would be right. All these answers are correct, but they vary in their granularity, how precise and constrained and non-ambiguous they are relative to the thing that you're trying to mention.

Or concreteness.

The meanings of words also vary in terms of how apprehensible the referents are. So you can talk about things that are very directly perceivable through your senses. Such [00:05:00] as rock, table, cell phone, things that you can touch, smell, or hear, right? But you can also refer to things that are very abstract.

You can talk about anxiety, about peace, about yearnings and longings, right? And these things, they are not so concrete, are they? So this is also measurable. And when we have to choose which words we want to use, we unconsciously have to make commitments between all these factors, and we can favor the words that are more easily retrievable.

Typically, the shorter ones, the more frequent ones, the less granular ones, or you can favor ones that are more challenging, that require a little bit of extra juice on your part, right? Namely, longer, less frequent, conceptually more precise words. And that's precisely what we are measuring here.

The point is that all too often, we [00:06:00] confound a task with the measures that we use to assess the task. People all too often say, "Oh, semantic fluency is very useful," or, "It's not useful." And that's the task. What they're really referring to is counting correct words, but that's not the only thing that you can count.

And when you look behind the obvious and you start finding other things that can be measured, you have this plethora of information or variables that can be useful to a number of things like phenotyping Parkinson's disease patients.

Dr. Michele Matarazzo: Fantastic. Thank you very much for guiding us through these different characteristics of language. I definitely will think about it in a different way when I ask my patients to tell me a list of words that start with a specific letter, for example. Now, you studied a large cohort of more than 460 people with Parkinson's disease, and remarkably, the input consisted of only three 1-minute [00:07:00] tasks: animals, supermarket items, and words beginning with the letter P.

So what were the main questions you wanted to answer with just these three minutes of language?

Dr. Adolfo Garcia: The first one is the one that I just mentioned, right? Which are the vocabulary search strategies or patterns that are most predictive of cognitive outcomes in a large Parkinson's cohort. A key feature of this study is that all patients were very well neuropsychologically characterized, and the first thing that we did was to see whether with the multivariate information that we extracted from each of these words, we could predict their overall cognitive status as measured by the Mattis Dementia Rating Scale, which is widely used and often recommended for cognitive assessments in Parkinson's disease.

So that was the first question. The second question was, we know that cognitive profiles vary substantially across the parkinsonian [00:08:00] population, in Parkinson's disease in particular, with different population-level studies suggesting that anywhere from 30% to close to 50% of patients, non-demented patients, may actually have mild cognitive impairment.

And the detection of mild cognitive impairment is quite challenging for a number of reasons. The second question we asked was whether these multivariate word properties would also allow us to classify between Parkinson's disease patients with and without mild cognitive impairment.

And then the third question was, once we do all this, which are the actual word-level properties, the actual word features that are most discriminative, that are most predictive of these cognitive outcomes, and that best differentiate between these two patient subgroups?

So those were the three guiding questions in our study.

Dr. Michele Matarazzo: Great. I think those are interesting questions. Let's get to the results then.

Dr. Adolfo Garcia: Absolutely. First thing we found is that you can train a [00:09:00] machine learning model with all these informations, the length of the words, the granularity of the words, the concreteness of the words, all these things. Instead of looking at them one by one, you can feed a machine learning classifier, train a model with all these features.

We left 20% of our patients outside this model. The model never saw them. They were completely separate from the model-training phase of the study. We saw that the model was quite good at learning from these word features how to predict the cognitive status scores of the patients.

And then once we trained the model, we tested it in these other patients that we left outside. And we found that we were able to predict their cognitive status significantly and with a strong correlation between their actual cognitive scores and the language-predicted cognitive scores.

This is interesting because of the [00:10:00] sample size. For some disorders out there, a sample size of 450 patients is not such a big deal, but for cognitive assessments in Parkinson's, it's not a trivial figure. And when it comes to training these models, sample size and feature-to-sample ratio matter a lot.

So we knew that we had the raw material to do something very robust in this sense.

The second thing that we found is that these machine learning models that we train were also very successful at discriminating between patients with and without mild cognitive impairment, with an area under the curve of over .75, if I remember correctly.

And all of this that I'm mentioning was done upon adjusting for sex, age, education, motor symptom severity as tapped by the UPDRS III subscale, and medication dosage captured through the levodopa equivalent daily dose measure. The final thing is that the main [00:11:00] drivers of these classifications, there were two or three features, but the one I will mention is one that we've developed and we refer to as semantic variability.

Semantic variability has to do with how semantically similar adjacent words are. So if you say dog, cat, backyard, for instance, right? The distance between dog and cat is quite short. The distance between cat and backyard is quite short.

But if someone were to say something like dog, hurricane, machine gun. The distance between those pairs of words is much broader, right? And something that we've found , and this is something that we found in an earlier paper, is that compared to controls, people with Parkinson's disease tend to have less semantic variability than controls. So, they tend to produce words that are closer in meaning to the previous words that they said.

And our reasoning behind this is that when you say a word, [00:12:00] if you need to jump to a different semantic space, if you want to go from dog to hurricane, right? One of the things that you need to do is to inhibit the active semantic field, the active meanings that are operative when you say dog, because to jump to something different, you need to inhibit what's already populating your mind, as it were.

And we found that in Parkinson's disease patients, this tendency to have lower semantic variability actually correlates with inhibitory symptoms. Not only that, it maps on the integrity of resting-state fMRI network that are putatively implicated in inhibitory processing. This is the one feature that was the main driver of the results that I mentioned in this paper.

Dr. Michele Matarazzo: Great. Again you're talking about very complex analysis of this quite simple data that you were starting from. But there is also an interesting comparison in the paper with the traditional approach of simply counting valid responses. So for [00:13:00] predicting overall cognitive performance, the number of words already contained a good amount of information, but when you try to distinguish patients with and without mild cognitive impairment, there was a difference between the two different approaches, right?

Dr. Adolfo Garcia: Exactly. Exactly. There was a difference, and what's interesting is that when you train these models, you have better generalizability onto unseen patients with these multivariate properties than you do with the typical word counts.

To some extent, the comparison is somewhat unfair because when you're going for the valid count measure, the typical measure, the classifier only has three values per participant: the number of words for animals, the number of words for letter P, and the number of words for the supermarket items.

Now, when we train our models, from each task we obtain a myriad of features, all the things that [00:14:00] I mentioned under distributional statistics. So that may be perceived as an unfair comparison, but I would say that on the contrary, this is part of the richness of looking at this. When you stop thinking of words as single monolithic things that can only be counted, and you realize that you can dissect them and extract all these types of information from them, you realize that verbal fluency measures have much more to offer than what we've traditionally demanded from them.

Dr. Michele Matarazzo: Good. Now, I would like to move from the scientific findings to the potential clinical applications of these findings. There is this attractive idea here that this could eventually become a very brief and scalable cognitive assessment. A patient speaks for three minutes, the language is analyzed automatically, and we obtain information about their cognitive phenotype at the moment that we are analyzing it.

How close are we to that scenario?

Dr. Adolfo Garcia: We are as close as we can [00:15:00] be. We've created an app, which is called Tell, and we founded a startup, also called Tell, which is basically a digital speech analysis tool that is our go-to, let's say, box for embedding all these measures that we develop in our research.

Tell is very simple. You can use it on any device, on a cellphone, on a laptop, on a tablet. It can be used face-to-face when you are testing a patient, Michele, or it can be used in teleneurology settings through video calls, or it can be used offline, and it can be self-administered by patients.

And basically, all you have to do is, you have prompts, which may be written instructions or pre-recorded instructions, or there may be a visual prompt for you to look at and react and, ultimately, speak. And you record yourself. And then this app, it will pre-process the audio, will extract all the audio features that we need when we need them.[00:16:00] 

It will transcribe the language that was spoken, and then it will extract all these properties that we are interested in from the transcripts of what was said. Tell is already in use by over 200 active examiners in over 15 countries in leading institutions worldwide, including Harvard Medical School, Imperial College London.

But the thing that makes us proud here is that our mission has to do with underserved communities and underrepresented regions. Tell is most particularly used in low- and middle-income countries, spanning Argentina, Chile, Colombia, Peru, Brazil, India, Africa.

We have a project running in Nairobi in Kenya. And we are already generating reports that summarize results and that are already being used to inform clinical decision-making and are being incorporated into patients' digital clinical histories. [00:17:00] So, the gap between what this could do and what can be done is being bridged, not just by us but by many other teams.

That's one of the beautiful things about this particular juncture in history where all these technological breakthroughs that a few years ago seemed more wishful thinking than concrete realities are actually making that transition.

Dr. Michele Matarazzo: That sounds fantastic. That brings me to a couple of other questions. One of them is, so you've been talking about the results on a group basis, but if we go to a patient basis, on an individual level, can we be so confident that the results that we are giving patients or they are giving clinicians to inform the patients are that accurate?

And then the other question is more of a technical one. Of course, language is the quality of the language, but apart from that, there is also the voice. Have you ever thought about combining information that we can get not just from what the patient is [00:18:00] saying, but also from how the patient or the subject is speaking? So, the voice.

Dr. Adolfo Garcia: Absolutely, yes. We have two worlds here. There's the language world, and there's the speech world, where speech is understood as the combination of motor physiological gestures that allow you to produce sounds, through perturbations of the airwaves. And then you have language, which is what you can extract from the transcribed text.

The content, the syntax, the affectivity of things and whatnot. Our first paper in Movement Disorders, in the journal, was published in 2021, and it was about actually motor speech measures in Parkinson's disease. So from the acoustic signal you have all these very detailed proxies of neuromuscular dysfunction, basically.

Hypokinetic dysarthria is one of the earliest, even preclinical, signs of Parkinson's disease, alongside [00:19:00] anosmia and other disorders. When you talk, you need to coordinate your muscles to produce all these subtle movements that result in speech, in intelligible speech.

And there are many things that we can measure. We can measure speech timing, for instance, how many pauses are left in between words, or even within words, and how long and variable those pauses are, for instance. We can measure voice quality, how discernible different segments in a recording are as speech versus non-speech.

We can measure pitch changes. One of the earliest manifestations of hypokinetic dysarthria in Parkinson's disease is mono-pitch. All of a sudden, patients start talking more monotonously without so many changes in their intonation patterns, and it has communicative values and it has functional implications, right?

So all these things can be objectively, automatically, and very swiftly derived from a digital acoustic signal, and the notion of the combination of information from both worlds is [00:20:00] fundamental because, in general, there are studies that look only at the acoustic signal, motor speech.

There are studies that look only at language, all these features that you can extract from transcripts. But in general, you have the highest discrimination rate, the highest prediction of neural dysfunction or cognitive outcomes and whatnot, when you actually mesh, when you combine information from both modalities.

And that is something that can be done nowadays very powerfully through machine learning. So we move from this single- variable analysis to this analysis where you can conflate all this information of multiple sources, and this leads me to your first question. Because another advantage of machine learning is that you move away from group-level results.

This group had significantly higher or significantly lower X than the other group, and you can start having probabilistic subject-level predictions. When I told you [00:21:00] that we left a number of patients outside of our trained model, we then can see how many of those patients are correctly classified as, for instance, Parkinson's disease with or without mild cognitive impairment.

So this is a step closer to one of the main requisites of precision medicine, which is individualized predictions rather than group-level estimations. Judging by this study that we are talking about, the 2021 study that I mentioned, and over 25 papers that we've published on the speech and language profile of Parkinson's disease patients with machine learning, I can tell you that these measures are variably, but overall consistently, robust at providing subject-level rather than group-level predictions.

Dr. Michele Matarazzo: Very interesting and definitely a lot of work to do in validating all these things and combining information, which is obviously exciting. And one other question I had, you were mentioning at the [00:22:00] beginning of the interview how proud you were of the Latin involvement of this study.

But inherent to this is that your participants were Spanish speakers. Of course, language is something that is variable from a place to another, even within the same language. Now, many of the features that you analyzed might depend on the structure and the statistics of the language.

Do you expect that the underlying cognitive signature, things like semantic variability, granularity, do you expect that they generalize across languages? Or will these models have to be developed independently for each language and cultural context? You were telling me already that you have data from this Tell app you're mentioning from very different countries.

So I guess you have an answer to that question already.

Dr. Adolfo Garcia: What a beautiful and timely question. One of the things that we really care about is this issue of cross-linguistic fairness, right? As you might expect, the bulk of evidence from speech and [00:23:00] language outcomes in not just Parkinson's, but neurodegeneration in general, it comes mainly from English speakers in high-income countries.

Over 65% or 75% of papers out there, they target these populations in particular. This is a paradox because when you take a look at how many people around the globe, including native speakers or second-language speakers or third-language speakers, actually use English, the figure is 17%.

This means that 83% of the world does not speak English or use English at all, and yet English accounts for roughly 65% to 70% of published studies on speech and language in Parkinson's and other brain disorders. That's inequity. That's Inequity 101. So one of our missions is to try to subvert this a little bit.

Another effort that I think is worth mentioning is the INCLUDE Network. It's the first international network for cross-linguistic [00:24:00] research on brain health. We created it 4 years ago, and now it's grown quite substantially. We have over 200 members from close to 100 sites, spanning 30 countries or so.

And the main things that we are doing now is compiling preexisting information, language data from all these sites, harmonizing it, and basically comparing it so that we can detect which of these language markers may be more cross-linguistically generalizable, and which ones may be more language-specific or typology-specific.

As to your specific question, Michele, these particular word properties that we are talking about here, granularity, semantic variability, my hypothesis is that they should be quite consistent across languages because they are not so much contingent on the structural, the idiosyncratic morphosyntactic or phonological, properties of the language.

They have to [00:25:00] do rather with the tendencies or the strategies or the constraints that patients have when they have to choose which words to access. And the tendency to favor more easily retrievable words, I would presume would be something rather shared across different languages.

Now, there are other aspects for which I would absolutely hypothesize major cross-linguistic differences. In fact, in 2020, we published this paper in Cortex where we assessed morphological patterns in spontaneous discourse from Parkinson's disease patients and controls who spoke Spanish from Colombia, German from Germany, and Czech from the Czech Republic.

We found that these morphological features, that is to say, how words are structured internally, that they gave you very good patient-control classification, and they were reliably predictive of UPDRS III scores, [00:26:00] but the specific morphological patterns that worked for each language were substantially different.

So yes, this is a point that really needs to be stressed because all too often we confound the results of one study targeting one language with evidence for how language at large may relate to a disorder, and that's a massive overstatement. It does not follow from the data, especially when you're targeting English, which is a very peculiar language relative to all the other 7,000 living languages in the world.

So this is vital, and I'm very happy that you bring this message up because we need to be more culturally and linguistically aware when we are designing, reporting, and then discussing these types of studies.

Dr. Michele Matarazzo: Great. I think that in general bringing up this issue is important. I think there's been, in the last few years, an effort to [00:27:00] try and have more representation and more diverse representation in research, but we are far from having an equitable representation so far. Let's see what happens in the future, and thank you for contributing to having a more diverse contribution to science, which I think is very important, especially when we talk about external validity of the data that we're talking about.

I also have another question because one thing I was thinking about when I was reading the paper is that when I am in the clinic with a patient, of course I would like to know from a very easy quantitative analysis whether the patient has or not mild cognitive impairment. That would be great.

But one thing that would be even more interesting is to predict whether the patient I have with me will develop, we're talking about cognition here, so will develop a feature of Parkinson's in the future or how likely the patient is to develop, let's say in this case, cognitive problems in the future.

Of course, this is a cross-sectional study, but do you think that [00:28:00] this might have a role in the longitudinal assessment of patients?

Dr. Adolfo Garcia: Absolutely. Absolutely. That's in general one of the beauties of digital tools. They lend themselves nicely to longitudinal studies. And not just in the sense that, okay, let's see what happens after 6 months or after a year, but even in much shorter time spans, right?

Because you can very easily manipulate the instructions or the stimuli that people have to respond to, and you can very easily circumvent things like familiarity effects or task-learning effects or washover effects would typically limit the usefulness of more traditional pre-packed tools for longitudinal testing.

There are antecedents of using digital speech and language measures in Parkinson's. There are some studies which are interesting in the sense that they've looked at these markers as preclinical markers. So you can find some of these alterations in [00:29:00] people with rapid eye movement sleep disorder, for instance, which is one of the main preclinical precursors of Parkinson's.

And you can find that the detection of these speech anomalies in people with that condition is then predictive of who is going to convert or who's going to develop Parkinson's, to be more precise. There are very few studies that have looked at longitudinal decay of these speech and language metrics.

More interesting ones are those that take a look at speech language measures in a particular time point, and then you try to predict ulterior measures in cohorts. There are very few. Like I said, we ourselves are working on our very first longitudinal studies right now.

This is so incipient that I don't even have preliminary results to share with you. I would love to, but it'll take us a year, I think, to have a paper published with these data. It's a beautiful cohort that we put together with colleagues from Brazil, and we have a nice initial cohort at [00:30:00] time point one, at T1, and then we have 70% of that cohort at T2, and then something like 50% that made it to a T3, with the time spans elapsing over a year.

So this is powerful. And the evidence is very incipient. But I think we're going to be seeing more and more of these studies as these automated tools become more and more popular.

Dr. Michele Matarazzo: Fantastic. Now, I wanted to ask you about the future. How do you plan to keep studying the language of Parkinson's disease, and how does this study and the future of this research fit in the global landscape of what you're doing about the language?

You're neurolinguistics, right? So you're not just studying Parkinson's disease. You're focusing on different disease and disorders. So how does this research fit into this landscape?

Dr. Adolfo Garcia: Yeah. This is one hub of a much larger ecosystem that we've developed. We've learned, like I [00:31:00] believe most of brain science over the last decade or so, to work in a very cooperative, multicentric, international fashion, and that's what we are doing. I think this is the way to go.

It's the way to increase representation of historically underrepresented communities. It's the way to reach larger sample sizes. It's the way to factor diversity into your designs rather than skewing it away as something to be matched or discarded as a nuisance variable.

Because it complicates things. The main things that we are doing now is first, we are aiming to build larger and larger cohorts for speech and language studies, in particular. But for neurocognitive studies on neurodegenerative disorders, underpower is a recurring issue.

That's one of the main things that we are working towards. This particular paper with close to 500 participants is a step in that direction.

The [00:32:00] other thing is multimodality. We want to see how these speech and language measures that we obtain, how predictive they are of alterations and relevant markers in other dimensions, including cognitive outcomes like we've done here, but also measures of motor symptomatology, measures of brain atrophy, or fMRI connectivity.

We also work integrating these measures with electrophysiological markers derived through EEG and biofluid markers. I think they are also fundamental. Perhaps not so much for Parkinson's as they are for Alzheimer's, nowadays, with the arrival of p-tau217 and other plasma-derived biomarkers.

But the multimodal biological validation of these markers is another key milestone to hit. And diversity. I will repeat myself, but trying to see which of these markers are most robust against multiple sources of [00:33:00] heterogeneity. So which of these markers are really robust when you think about their variation based on country, sex, level of income, motor symptomatology, the medication that the patients are in, and whatnot.

I think that's fundamental because going back to the notion of personalized precision medicine, if we keep doing science where we compare carefully crafted groups at the group level, after we've excluded all participants who are tricky because of their age, because of their comorbidities and whatnot, then we're not really doing much for the actual patient with comorbidities that actually shows up at your clinic, Michele, because that's the reality of things.

There's a divorce often between the pristine nature of scientific designs and the highly heterogeneous, complex realities of the individual patients that you and neurologists have to work with. I think [00:34:00] that bringing diversity into the designs and trying to find the factors that are most robust despite all that diversity is one of the key changes in mindset that we need to embrace, that we are working towards, and that will be necessary for science to be less about idealized group-level differences and more about personalized, clinically actionable information.

Dr. Michele Matarazzo: Fantastic. And finally, Adolfo, this paper has now been selected as the Best Research Article of the Year in Movement Disorders. Looking beyond the individual finding of the study, what do you hope the movement disorders community takes away from this work?

Dr. Adolfo Garcia: Speech and language matter. Speech and language matter. They matter a lot. I remember reading a very popular review article, I believe it was in Nature Reviews Neurology, on overall characterization of Parkinson's disease. I won't mention the [00:35:00] authors, but it was published in 2009, and language was dismissed with just one sentence saying, "Language is typically preserved."

That's all that was said about language in Parkinson's in this very, very high-profile review paper in one of the top journals. 2009! I'm not talking about 1960 -- 2009, right? I think that we've come, I don't know if a long way, we've come some way from that now in 2026, but not enough. I would like the field to become more attuned to the contributions that speech and language measures can bring to the table.

Not to replace anything, not to substitute anything, but to add to that joint transdisciplinary effort. There's so much information, so much signal, about motor disorders in speech and about cognitive dysfunction in language that I think that we are doing a disservice to the field and to patients if we skew this [00:36:00] source of insights just because of traditions and acquired habits.

That would be the main takeaway. Let's pay more attention to language. Not because it matters more than other things, but because the opposite is equally false.

Dr. Michele Matarazzo: I'm just remembering when I was a neurology resident a few years ago, one of the neurologists I was training with, a very good neurologist but who was a senior and he was not a movement disorder person, he was doing pediatric neurology, saw all this world of non-motor manifestations of Parkinson's. He was very excited, and he told me one day, "This is like Parkinson's without Parkinson's." And I think you're right. We've come a long way from those early days and definitely there is still a lot to study and understand.

Hopefully this will help us from a clinical perspective, from a pathophysiological perspective. So Adolfo, thank you very much for being with me today. Congratulations again to you and to the entire team on this recognition. And thank you for [00:37:00] this conversation we've had.

Dr. Adolfo Garcia: Thank you very much. Props again to first author, Franco Ferrante, who led this beautiful paper, to the entire team that made it possible, and thank you again, Michele, for this space. It's so valuable for us to have a chance to voice what we are learning about voice in Parkinson's, and what better place to do it than here.

Thank you very much once again.

Dr. Michele Matarazzo: And to everyone listening, thank you for joining us on the MDS Podcast. [00:38:00]