Can Your Voice Reveal Early Signs of Alzheimer’s?

Jump to Transcript

In this episode, Lauren & Matt speak with Neguine Rezaii, MD, an Assistant Professor at Neurology at Harvard Medical School and a Board-Certified Neuropsychiatrist at Mass General Hospital. Dr. Rezaii’s research focuses on using natural language processing and artificial intelligence to improve the detection and classification of neurodegenerative disorders.

Dr. Rezaii talks about her recent study, Voiceprints of Cognitive Impairment: Analyzing Digital Voice for Early Detection of Alzheimer’s and Related Dementias, “which explores how artificial intelligence can analyze short speech samples to identify subtle signs of cognitive impairment. The study used speech recordings from 120 individuals and found that features such as pauses, word choice and the specificity of language can help distinguish cognitive impairment and Alzheimer’s disease. 

The conversation also explores how a simple 30-second speech test could potentially be used in clinical settings to support earlier detection and monitoring of cognitive impairment, as well as the challenges and ethical considerations that come with using AI to analyze something as personal as a person’s voice.

Article Discussed in the Episode: 

Rezaii, N., Wong, B., Aisen, P. et al. Voiceprints of cognitive impairment: analyzing digital voice for early detection of Alzheimer’s and related dementias. npj Dement. 1, 35 (2025). https://doi.org/10.1038/s44400-025-00040-0 DOI; npj Dementia volume 1, Article number: 35 (2025)

Transcript

Matt Davis:

Imagine if the way that you retell a simple story could reveal the earliest signs of dementia, long before symptoms become obvious. In this episode, we're discussing a study that examined how artificial intelligence can be used to analyze digital voice recordings to detect cognitive impairment. By picking up on subtle linguistic and acoustic markers that the human ear often misses, this noninvasive tool can identify early signs of impairment with surprising accuracy. Today, we'll explore what this may mean for the future of early detection of dementia. I'm Matt Davis.

Lauren Gerlach:

I'm Lauren Gerlach.

Matt Davis:

And you're listening to Minding Memory. Today, we're joined by Dr. Neguine Rezaii. Dr. Rezaii is an Assistant Professor of Neurology at Harvard Medical School and a Board-Certified Neuropsychiatrist at Mass General Hospital. Her research focuses on leveraging natural language processing and artificial intelligence to improve the detection and classification of neurodegenerative disorders. Her work has investigated primary progressive aphasia and the use of large language models to identify natural categories of ekphrastic subtypes from speech samples. Additionally, Dr. Rezaii has been studying early language impairments among patients with cognitive impairment. Neguine, welcome to the podcast.

Neguine Rezaii:

Thank you so much. It's a pleasure to be here.

Matt Davis:

Dr. Rezaii was the lead author of the study titled Voiceprints of cognitive impairment: analyzing digital voice for early detection of Alzheimer's and related dementias that was published in the npj Dementia. The study used artificial intelligence to classify speech samples from 120 individuals to detect cognitive impairment. We'll include a link to the study attached to this episode. So, before we jump in and talk about your paper, I was wondering if you could start off by telling us how you got involved in this area of work. We're interested in what drew you to neurology and how that interest evolved into studying speech.

Neguine Rezaii:

Sure. So, with this particular question, I would be a little bit broad for at least two reasons. One is because it actually is the way that things unfolded in my life. And then secondly, I'm going to borrow some of the concepts that I'm going to use in the introduction, so it's going to be slightly asymmetrically broad here. So, many people are fascinated with fundamental questions. By that I mean fundamental questions in philosophy. For example, in metaphysics, we've got this question of what is being. Why do we have something at all instead of nothing? Because that's a possibility to have nothing at all. And some questions are more in the domain of epistemology. How do we know something? My quest was slightly more in the domain of philosophy of perception, to what extent things that we can know are reflective of an outside reality. Is there an outside reality? And how do we relate to that?

So, these questions engaged me for a long time. And I'm sure for many people it is the case. It is just the matter of the extent of engagement. This is my recent observation from my acquaintances. But those people who go to physics, not all of them, but many, at least in my life, are individuals who still have this idealism about learning more about reality, about being. So, they go to physics with a hope of learning more about this metaphysical question. But soon, there is this moment of disappointment when they find themselves getting the second derivative of something with respect to something else. And then they realize that this is not what they were after. That moment of disappointment also happened to me, but a little bit differently. So, I was very much interested in thought.

What is thought? Is it a composite entity? Is it unimodal? Can we break it down? What are the materials of thought? So, these were the questions that engaged me. And I knew very soon that it's going to be neuroscience, a kind of a neuroscience. Either neurology, psychiatry or just pure neuroscience. So, I have different methods very early on. For example, one question was that, how can we know something? Soon you realize that our perception is not the most reliable way of knowing things. Anybody who wakes up from an immersive dream can tell you that. We all have these experience of being very much engaged with dream imageries, like a woman, a tall woman, long dress, long black dress, comes in nonchalantly with a hat in the form of a saxophone. You engage and you don't even question it. And you wake up, you realize that, wow, it was all a dream. And as Descartes says, "How can I ever trust the reality outside if I cannot trust my own senses?" He talks about this example from dreams.

So, for a long time, I was just logging down my dreams. I have thousands of them. And I was also logging my mental distraction. For example, in a psychology lecture, I just found myself thinking about an earthquake in my childhood. And I was, why am I thinking about this? I realized that a lecturer was talking about Rorschach test that I thought about its phonological associate, which is foreshock. And then I thought about earthquake and the early earthquake that I had in my life. But the problem with all of these was that they were very subjective. That was the main problem. So, that was my moment, my moment of disappointment. I found language as a fascinating proxy. So, I was lucky that I found this alternative way of approaching something that really fascinated me. So, language is objective. You do not have to rely on people's personal logs.

Language is measurable, even more so thanks to all these AI methods, natural language processing methods, language models that are available. They're even more measurable these days. It is composite. We're going to talk more about that. Each aspect of language can tell you something about your cognition, not just cognition. For example, when you say like, "I have stomach pain." Basically what is happening is that the pain receptors in your stomach are going up the spinal cord to the brainstem and all the somatosensory areas. And then from there to your language network and you say, "I've got pain." So, you can argue that language has access to what's going on in your stomach. And it also has access to your subconsciousness, which is fascinating. When you study, for example, Freudian slips, it's all about what you say that you do not want to say. And you can study them objectively.

So, these are all the reasons that I found language as a fascinating proxy to study thought. And thought can break down in different ways, in different neuropsychiatric conditions. During my studies, I studied patients with schizophrenia, very severe psychosis, hallucination, delusions, and later on with neurodegenerative disorders. This is a long answer to your question, but I will be borrowing a lot of concepts from this introduction.

Matt Davis:

No, that was great. I was struck by just how much science and questions and informed your career path. You're a physician, so where was the clinical interest in terms of your decision tree, in terms of what you pursued over perhaps a PhD?

Neguine Rezaii:

The breakdown of this. So, there are different ways of knowing a system. For example, if there is a clock that you've never seen, you can just take the parts out and see its impact on the way the clock works. So, this happens in the clinical realm. When a part of the brain is damaged for a reason, then you see its impact on, for example, in this case, language. And you may wonder, maybe that region is responsible for that particular task. So, in addition to neuroscience, I found that the clinical domain gives you access to another dimension of how something works, which is by lesion studies.

Matt Davis:

And over your trajectory, I guess, where did you pick up some of the knowledge that you have around artificial intelligence and machine learning and those types of things? Were those things that you acquired throughout your experiences and research and collaborations?

Neguine Rezaii:

Exactly. So, I'm coming from purely medical background, but I always like math. And in that quest that I shared with you, I was also thinking about going into just studying mathematics. I thought there are some answers there, but we have to choose only one major, it seems like. But I remained very interested. So, I went to some classes. There were so many open course lectures outside that you can just participate in. A lot of online videos that you can just see, which helped me a lot actually.

And through collaboration, I learned a lot and I really enjoyed doing the programming myself. These days, it's a lot different, a lot easier. Debugging a program used to be very challenging and fun, but these days you just ask a large language model to debug it for you, which as researchers recently found it will cause some mental bankruptcy. This the term that they've used and I agree, but it's faster and more efficient.

Matt Davis:

So, turning to speech, I'm curious. So, are subtle changes in speech something that you personally have noticed in the treatment of your patients with early onset of cognitive issues? Is that something that you noticed originally in the clinic?

Neguine Rezaii:

The honest answer is no. It did not start that way. Even though when I got sensitive about it, then I start to pick up on subtle changes a lot better. But the trajectory was that I just wanted to study thought and use language as a proxy. And then I found that there are different ways of breakdown and each way can give you different insight into neuropsychiatric disorders. But once you realize that this is the problem, there is no way that you cannot see it anymore. We're going to talk more about the informativeness. The breakdown of being informative is something that happens in Alzheimer's disease. And once you know that, then you become sensitive.

I always tell my friends and colleagues, just listen carefully to people's language, to your loved ones, to your relatives when they're aging. Are they becoming more vague in their language? Is it the case that they talk to you and then after the conversation, you have no idea what they talked about? Maybe it is early Alzheimer's disease. Just be attentive to these little changes. So, the sequence was from recognizing some patterns and then seeing it in reality.

Matt Davis:

In broad strokes, I mean, some of our listeners have clinical backgrounds and some don't. I guess I'm curious, do neurodegenerative disorders affect speech and language in very different ways?

Neguine Rezaii:

Exactly. And that's the beauty of it. I don't know whether it is by chance, but we are so lucky that language is composite. It is not a unimodal entity. When we talk and we don't even think about it. It's not even effortful for us. We just talk. For example, T and D, they're placed these two phonemes. The placement in our mouth is very close, but we don't even think about it. Where do I put my tongue in my mouth to say T versus D, even though they're very close? So, it is very effortless. We don't even think about it. But there are so many networks that are engaged. One is processing memory. What happened? The thing that you're going to talk about, what was exactly that thing? Or for example, if you're describing something that you saw, the visual network is active.

Then you need to get the right words. There's a semantic network mostly behind our ear, left ear. There's a region there for getting all the words and the actual concepts. And then there's a region, there's a lot of disagreement upon where exactly in the brain is responsible for syntax, putting all these elements together. An entirely different network is involved. And then optimize information. Make sure that you're being informative, what you're saying is actually making sense. Also, adjust it to the emotional status of the person you're talking to. Is the person sad? Maybe your tone should be different. So, there are so many networks of different types that get active. And then everything is combined at the end and you've got a speech sample. And if you can break it down, then you can see different cognitive functions. As each network breaks down, you will see a different impact.

And of course, there's a whole topic about compensation. If we are not able to say things in words because the brain region behind the left ear is damaged, what should I do? I don't want to become mute. I still want to talk. I still want to share my experience. If I have that lesion, are there ways to compensate for it? And in one of our papers in 2022, we found that, yes, you can. If you cannot say things using words, you can use more complex syntax. For example, instead of saying sailboat, if the sailboat region is gone, you can say a boat moved by the wind. It's a longer sentence. Words are simpler. Syntax is more complicated, but you still can convey your message.

Lauren Gerlach:

So, let's jump into your study now. Can you tell us a little bit about what you did and what type of data you used?

Neguine Rezaii:

Sure. This study, I cannot talk about study without talking about the senior author, Dr. Brad Dickerson. He has been my mentor since the fellowship in Harvard. And he's been the best mentor that I could ever have. I always have the liberty of choosing the topic and be as exploratory as I wanted to be. So, he had this data. It is called LEADS, Longitudinal Early-Onset Alzheimer's Disease. So, it's for individuals who get the Alzheimer's disease or cognitive impairment, to be more accurate in this case, who get the disease before the age of 65. The diagnosis in these individuals are a lot more challenging, because if somebody's 47 years old... one of my patient is that old. When somebody's 47 years old and they're not being very informative, everybody thinks about maybe it is depression, maybe it is menopause, maybe it is sleep deprivation, maybe it was COVID.

So, people rarely ever think about Alzheimer's disease at that age, so this is a particularly challenging diagnosis at that age range. So, I find this maybe even more helpful to have the opportunity of working on this data set, because then you can help individuals who have some cognitive impairment, but they're not related to all of these factors that I shared. It is maybe Alzheimer's disease. So, this work is based on the LEADS data, language samples, patients, and it also has healthy individuals. So, the participants were asked to listen to a story, a very short story, about 20 seconds, and then repeat it immediately after, and then 20 minutes after. And then we analyze those language samples.

Lauren Gerlach:

And it sounds like this unique cohort is collecting a lot of different data. Can you tell us a little bit about the types of things that they're measuring through this cohort?

Neguine Rezaii:

Yes. So, it is an amazing cohort because it has objective data. Early work, we have this fantastic DementiaBank. This also has a large cohort of speech samples from patients with Alzheimer's disease. But the wonderful thing about this LEADS data set is that it has biomarkers, objective ways of detecting Alzheimer's disease to confirm. Because in the DementiaBank, because it was a very early work and it is impressive for a number of reasons, they started very early. But at that time, biomarkers were not available. Now they are available, so we can confirm for sure that these individuals have Alzheimer's, because even the best clinicians can be wrong. We can be about 80, 85% certain about the diagnosis of Alzheimer's disease just clinically. So, we do need additional information, MRI, getting some samples from the CSF, cerebral spinal fluid, to measure certain proteins that can tell us for sure whether it is Alzheimer's disease or not.

Or amyloid PET scan, look for Alzheimer's plaques in the brain through a PET scan. And one of the strengths of this cohort is that it has all these biomarkers available. It is important so that when you develop a machine learning model, you know that the labels are based on ground truth. So, you do not have uncertainty about is it really Alzheimer's or not. So, this is what is unique about it. And in addition to this most important factor, biomarkers, patients have lots of neuropsychological testing. And also, equally important, it is a longitudinal study. So, they follow patients for years and years.

Lauren Gerlach:

Great. And it sounds like as part of that, battery participants were asked to recall a brief story and then describe at various time points what they could recall from it. For those listeners who may not be familiar with the Craft Story Recall task, can you describe just briefly what participants are asked to do and how that works?

Neguine Rezaii:

Sure. So, the story is very short. Lauren, you can test my Craft Story Recall task now. It was based on 2024 work, so hopefully you give me that... You're more-

Lauren Gerlach:

I'll give you some leeway there.

Neguine Rezaii:

Thank you. I appreciate that. So, the story is about Maria's child, Ricky, who loved to play soccer every Monday at 3:30 PM. One day, he kicked the ball so hard that it went over the fence where three large dog lived. The owner of the neighbor came out and gave the ball back to the boy. I probably get like, I don't know, 70% of the... but it is something like that. A very short story. And people are asked to say it immediately after and then 20 minutes after. And this has been a very, very helpful test. Analyzing speech sample there for different reasons that we're going to talk about more, it is very helpful.

That does not necessarily mean that other language samples would not be as helpful. We have found picture description tasks, like just look at this picture and tell us what is going on in that picture. We have found that also very helpful. But the design of this test is because it relies on memory. It is more suggestive of episodic memory functioning.

Matt Davis:

That's what I was going to ask. It does seem like the recall is an important component rather than just reading the story. You're looking for the ability of the mind to recall what was said previously as a component of the speech patterns that you're looking for, right?

Neguine Rezaii:

Exactly. And one of the most classic symptoms of Alzheimer's is amnesia, difficulty recalling things. So, that is the logic behind designing this test. But again, we have found picture description task as helpful with similar accuracies. So, we're going to talk about why it is the case, why there are some features in language, that independent of a specific type of task, they still give you clues about the diagnosis.

Lauren Gerlach:

And without getting too technical, can you tell us a little bit about the machine learning and data analysis that you use to analyze these voice recordings?

Neguine Rezaii:

Absolutely. So, we had two different approaches. One of them we call glass box or white box. I mean, this is a term that exists. It's not something that we coined, but it is to show the contrast with the concept of black box. What do I mean by that? So, in the first approach, glass box, we extract certain features, very specific features. For example, how many words did I say in a minute? Just measure it. Easy to measure. These days it's very easy to measure speech rate. Or tell me, for example, the ups and downs in my voice. Easy to measure these days. Tell me how complex the vocabulary was. Tell me how close the vocabulary that I used, where to the vocabulary in the actual story. So, these are very specific features. There are about 200 of them and we extract them from each sample.

And then there's some classifiers. There are different types of classifiers these days that we use and say, okay, this is the language. We tell them in advance. We tell the model in advance. That's why it is called supervised. We tell it, okay, this is what the language of patients with Alzheimer's look like and this is the language of somebody who does not have Alzheimer's. And then the model learns. It gets trained to look for features that are very diagnostic of Alzheimer's versus those that are not. So, it selects certain features in the training process that, okay, I was very wrong in the first round of classification because I chose the wrong feature. Let us try another feature. Oh, this time my performance is a lot better when I compare to the ground truth. So, maybe that feature is important. So, because we know exactly what features went into the classifier, it is called glass box.

And then at the end you identify the features that were most helpful. There is another method that we call black box. It is actually more... the performance is higher for the second approach. There's always in life... This is life, right? There's always a trade-off. So, if you get something better performance, you lose something. And in this case, it is transparency. So, you know that it's going to work better. You just get a language model. It can be GPT or there's so many language models these days. DeepSeek, for example. You just get one of them and you don't tell the model anything. You do not give any features. You just say, okay, look, I have 100 speech samples from patients with Alzheimer's and I've got 100 speech samples from patients with non-Alzheimer's. I just tell you the labels. You search in that and find whatever feature that is most helpful in order to do the classification.

When we say here, feature is not the best term to use because it is not as explicit as speech rate, because it could be just a very bizarre combination of a word in the context of other words or in combination with an acoustic feature. So, it's like a statistical pattern rather than like a very explicit feature. So, we don't know exactly what went into the model, but they tend to perform better slightly, like 90% versus 95% when we compare the two methods. But now there are some ways... They call it XAI or explainable AI. Some methods to try to find out what helped in the classification.

We're trying to find out, get some insight from the black box models. They're not perfect, but at least it gives you some idea. So, these are the two broad measures. One of them was feature based and then the other one, "You, language model, do whatever you want. But at the end, what I expect from you is accurate prediction."

Lauren Gerlach:

And the types of things they might pick up on would be like paucity of speech or maybe adding in features that were not part of the original story, or are these some of the things or pauses just like big picture, what would be some of the differences that they might be picking up?

Neguine Rezaii:

In the first or the second approach?

Lauren Gerlach:

Actually both, because it sounds like the first you're maybe going in with more of a priority.

Neguine Rezaii:

Yes. In the first, things that we found out that were helpful was pause duration, how much I pause between words. Maybe because I cannot think of the words. Maybe just retrieving the right word is very difficult. So, people tend to pause a lot. And then specificity, this is very important in Alzheimer's disease. Patients with Alzheimer's, I'm just talking about pathology proven Alzheimer's and when we have the actual biomarkers. These patients are non-specific. They can tell you she went there. You wonder who's she? Where? So, they use words, maybe syntactically. She went there, it is correct in terms of syntax, but you don't know who she is. So, they use a lot of pronouns. You do not know where there is. So, in terms of time, in terms of place and names of people, they are very vague, patients with Alzheimer's disease. And this is something that is kind of unique to patients with Alzheimer's versus other types of dementia or other reasons for cognitive impairment.

Somebody can be experiencing a brain infection or somebody can be experiencing very deep depression. They might have a language that is different from norm, but they might still be accurate and specific, but they might be slow. There might be a lot less variation in their tone, so they could be very monotonous in depression, for example, but they can be still specific. So, this loss of informativeness, this loss of being specific is something that is unique in Alzheimer's disease. And in the second approach, we kind of found the same thing. We only could go by text. So, with the second approach, we couldn't use the... because the large language models that we used was just text-based, not audio based.

These days, there are methods to work on audios. So, it was just the words. What words helped the classification the most? For example, we found that remember, remember, remember was a word that was helpful in the classification. In the story, we don't have the word remember. Why it is diagnostic, because in most of the speech samples, patients just say, "I cannot remember the name." So, that's why remember was something that was diagnostic. So, you can just find at word level, what were the things that were helpful. And we found that, for example, in the second approach, 3:30 PM, they're not using it. Specifier of time, they don't use it. Ricky, they don't use it. Proper noun, they don't use it. So, they were kind of converging. Both methods were converging towards informativeness in Alzheimer's disease.

Matt Davis:

It sounds like this was conducted as a side-by-side comparison of the two groups. I'm curious, have you started to think about the application within person? So, a big change for one person could be very different, so analyzing text across time.

Neguine Rezaii:

That's an amazing question. Yeah, this is wonderful. Right now, we're focusing on this question mostly to predict from baseline speech sample, how the trajectory of the disease will look like. Because once we make the diagnosis for Alzheimer's, especially in young individuals, people before the age of 65, like somebody in their fifties, they wonder what's going to be next, how my life is going to look like. So, this is usually an immediate question after the diagnosis. And now we are focusing on, can we predict who will have a fast progression versus slow, based on their speech samples at baseline? What are the features that can tell us about the prognosis of the disease? So, we're working on it. Hopefully, I will have the answers soon, but this is a very good question.

Lauren Gerlach:

That's a really important clinical applicability of this, because as a clinician, it's very hard to be able to answer those questions because every patient's trajectory can look very different and their risk factors and things that could contribute can vary. So, having some data or sense of what that might look like for a person, I think could provide very meaningful information to patients and their care partners. So, can you tell us overall, what did you find?

Neguine Rezaii:

The study had two steps. And this is based on the proposal from Alzheimer's Association, that if somebody comes into your clinic, they have some complaints about cognitive performance, then how you should start the process. So, the first step is, is there cognitive impairment or not? So, the first step was that classification. LEADS data set is strong for yet another reason, which is that it has three groups of individuals, healthy individuals, those who have cognitive impairment, not because of Alzheimer's disease, and those who have cognitive impairment because of Alzheimer's disease. Those who have cognitive impairment, but not Alzheimer's, are the ones that it could have anything. It could be depression. It could be midlife crisis. It could be menopause. It could be COVID. It could be a kind of infection, something. It's just not Alzheimer's, but it is cognitive impairment because of their performance, cognitive performance.

So, first, based on the guideline from Alzheimer's Association, the question is, is it cognitive impairment or not? And we found that in these individuals who have cognitive impairment, something that stood out was actually the pause duration. We talked about it. So, pause duration is an important factor, regardless of the specific type of pathology. And then the other factor that we found very helpful in that classification was number of sentences. Counterintuitively, this time, patients with cognitive impairment use more sentences, not because they're more informative, because their sentences are so short that they have to kind of compensate. They do not have the cognitive capacity to make a long sentence. So, in order to convey their message, this relates to the idea of compensation that we talked earlier. So, they use more sentences. So, their speech sample will look like longer, but shorter sentences. So, these individuals have longer sentences.

And each word, if you look at it, each word is very short. So, sentences are short. The components of the sentence which are words, they are also short. In order to compensate, they talk a little bit more in terms of sentences. The number of sentences increases. So, this was the first classification whether a patient has cognitive impairment or not. But for the second question, which is a harder question, is okay, two cohorts of cognitive impairment. One is because of Alzheimer's pathology and the other one is not. So, this is a much harder question. And for that reason, the performance goes down because it's a more difficult question. Is that, okay, who has Alzheimer's disease? This is the part when all these informativeness measures come in, how specific they have been.

What is the ratio of she, he versus proper names, Maria, Ricky? What is the proportion of that? Instead of saying he went to the field, to the park, to school, do they say here or there? So, how specific the adverbs are. How specific they are in terms of time, 3:30 or just say then. So, being very specific in these domains was the main factor, identifying a patient with Alzheimer's disease or pathology.

Lauren Gerlach:

And how well did the AI models compare to say more traditional measures, like the Montreal Cognitive Assessment or the manually scored story recall?

Neguine Rezaii:

So, one thing, well, if you just focus on accuracy, for the first step, MoCA, Montreal Cognitive Assessment test was not bad, comparable, because it's an easy task overall. AI was 99% accurate. MoCA was 92%. And so, comparable. But with the second question, is the cognitive impairment due to Alzheimer's, there we found the big difference between the two approaches. So, AI was still 85, 90%, and then MoCA suddenly dropped down to 60%. 59 to be very accurate. So, it suddenly dropped down because it just could not find the pathology. It did not have the... enough. It was not able to capture all these nuances that we talked about. One problem is that it's very categorical. If a patient cannot recall a word, should we measure how specific that word was? MoCA doesn't tell you that. It just tells you if you cannot say camel versus snake, versus rhinoceros, these are very different words in terms of how specific they are.

You will get the same score regardless of how specific that word is. So, that's why MoCA does not have enough sensitivity to capture AD Alzheimer's pathology. But that is only just comparing performance. We also need to pay attention to other aspects of a good test. It is not just about performance. Yes, that is perhaps the most important thing, but in real life, MoCA can take about 10 to 15 minutes. It can be associated with a lot of frustration, especially with patients who are really struggling with memory. And let's face it, clinics are already very much overwhelmed, so another test can be really difficult for both patients and the clinician. But this test, the Craft Story is only 30 seconds. So, it is saving you a lot of time by having superior performance. So, I think it is also important now that when we compare tests, we take into account other factors besides performance. Time, effort, how emotionally burdensome the test is for the patient.

Matt Davis:

Was there anything that surprised you in terms of the results that you found?

Neguine Rezaii:

Yes. So, this informativeness, now I need to be slightly broad here again. At least the way I view it, the goal of many domains in science is to find out the core driver. I remember when I was reading a book, this fascinating book, Evolutionary Psychology by David Buss, that it kind of modeled the world the way we choose things. Our taste, our sense of beauty, aesthetics. For the most part, they rely on survival and reproduction. And that was deflating to find that, okay, so I like this or I find this very beautiful just because of survival. I find this pastry very tasty just because it's in terms of calorie density, it is the most helpful in terms of survival, or I find this person attractive just because of reproduction. So, there is this moment of deflation when you find that the core driver for many humans' behavior, not just human, are these two factors.

Here, with language, we also need to pay attention to the fact that what is the core driver of our word choices or the way we talk. The core is, at least the way I'm viewing it right now based on the results of this, is conveying information. However, maybe in Alzheimer's disease, maybe it is unique in the sense that the core driver of communication is broken. In lots of other types of neurodegenerative disorders, for example, in aphasia disorders, patients can compensate. The example that I gave you early on, sailboat versus the boat moved by the wind. This is exactly what we found in patients with aphasia. They're able to compensate. If you cannot say a word, fine. Make it more complicated in terms of syntax and say it. Alzheimer's patients cannot do that. So, they just lose the core driver of language being informative. And it seems like they don't even know that.

So, in many textbooks, in neurolinguistics, they tell you that there is a center in the brain. By center, I'm being maybe metaphorical here. Maybe it's not specifically a region in a bit. Maybe there is, I don't know. That monitors the amount of information that you're conveying. Maybe in Alzheimer's, that region or that network is broken. And I was surprised by the extent of it and lack of insight. Again, patients do not know. They talk to you. They can talk to you for an hour. It's not necessarily paucity of speech. They talk to you, but they're just not informative. So, the extent of this breakdown and also losing the core driver of communication was surprising to me.

Lauren Gerlach:

Yeah. So, just thinking about the clinical applications of this, how far off do you think we are to using these types of approaches? And how might you see this technology best being implemented, thinking about supplementing in clinic versus home? Where do you see us going with this?

Neguine Rezaii:

I'm using it. I'm using it in my clinic. We have got a research study that we get verbal consent from patients. We record their voices and we just tell them. At the end of the clinic, I tell them, okay, this was your speech rate. This was the amount of informativeness. So, I use it and I love it. Patients love it. So, we've received very positive feedback that, I just get a test based on just 30 second. The score is based on 30 second tests. So, we are all excited by it. I mean, we're not that far really. Again, we are using it in our clinic. But one thing that is important is that to replicate it across studies that are very different. So, one of the strengths of LEADS is that... yet another, is that it is multi-site. There's so many. I believe 19, I need to double check the number, but 19 sites as part of LEADS.

So, it is a multi-site national study. So, you get the diversity that you need. However, it is important to see it in completely different population. See it in other languages. I worked with patients, Persian-speaking patients, and we found similar findings. So, one of the reasons that I chose Persian, one is, of course, I'm native in Persian. But the other one is that in the Indo-European branches, Persian is very far from English. So, it will be very interesting to see to what extent you replicate. Do the findings depend on the structure of a language or is it at least partially independent? We found that informativeness is still the core deficit in patients, in Persian-speaking Alzheimer's patients. So, that was fun to find out. So, we need these tests to be replicated. Larger populations, different languages. And then in terms of the technology, it's very, very easy and lightweight, I can say.

It can be very simple. What I have is just like a webpage that you record, you do some light computational methods, and then at the end you've got the scores. And it can be a very easy app that can be used at primary care setting. The examples that I used is that for heart condition, you have these smart watches. Or when you go to a primary care physician, you get your vital signs checked. But for brain, I mean, my other colleagues from other medical fields will disagree with it, but I'm going to say it. Brain, maybe the most important organ in the body, we do not have these methods. We do not have a vital sign for brain health. Hopefully, this 30-second quick test can become a way of a vital sign for brain health.

Matt Davis:

I mean, an important point is that your study investigated this among adults. Do you think the findings would differ much if you looked at older adults in terms of the speech differences?

Neguine Rezaii:

You exactly ask questions if you're working on it, so interesting to me. So, we are right now working on a cohort of individuals with the same pathology proven Alzheimer's disease, but much older. Hopefully, we will have answers for that question soon as well, but it is very important because these individuals were young. My guess is that it's going to be comparable, the results, but we won't know until we finish the analysis.

Matt Davis:

We're going to have to have you back sometime. So, there is a growing field, as you know, around using voice to detect and monitor mood, emotion, and psychiatric symptoms. I'm curious, from your perspective, how does cognitive voice analysis perhaps overlap and potentially differ from that of sort of mood and emotion detection?

Neguine Rezaii:

Sure. Yes. That's why having an entity that is composite is important, at least 200 features. And they are differentially affected in different neuropsychiatric conditions. In one of our papers in 2025, we work with patients with FND, functional neurological disease. These are individuals who can have similar presentation. They come to you and they provide maybe little language or that their speech is very different from... It just feels different. But when you analyze it, you find out, with this condition or a lot of affective mood disorders, you find out that it is more about the acoustic features. For example, the ups and downs in their intonation. Or for example, the variability in their pitch. You see a lot of changes there. So, variability in amplitude was the feature that stood out in the FND patients, functional neurological disease, but it absolutely was not helpful in diagnosing patients with Alzheimer's disease.

So, in Alzheimer's disease, it was more about the content of language, what you say. In FND was how you say it. I did work on the language of patients with schizophrenia, but not depression. But just reading the literature, a lot of it is about the way people say something, rather than how and the content of language. So, again, having dimensions help you differentiate these conditions.

Lauren Gerlach:

And as we think about wider-spread use of these types of technology, I think with some of the things maybe around monitoring mood and mental health symptoms, sometimes they talk about passive or repeated voice monitoring versus what you're talking about, a recall story. Just as we think about ethical considerations or how these things might be used moving forward, what kinds of things do you think the field should be thinking about?

Neguine Rezaii:

Yeah, this is tricky. This is very tricky. It is not like a DNA sample that needs permission from you, even though people still can get some cells. If there's some traces, they can do that. But we talk freely. It's available everywhere. Imagine a condition that a young individual goes to find a job and the potential employer runs an app and says, okay, this individual is going to get schizophrenia in two years. Would that individual hire the person? What are the ethical considerations of that? Are they allowed to run a speech sample? If it's going to predict.... The very famous nun study from JAMA 1997 showed that language can be predictive of Alzheimer's disease even in... It was a study on nuns who wrote down their diaries, so they followed them for 60 years and they found that, okay, the diaries in their twenties can be predictive.

So, what are the considerations and all the consequences of? We just talk freely and people can just run an app. Okay. This is Neguine's fate. She's going to get Alzheimer's disease with this chance, with this probability at this age, this disease, depression, with that probability. So, you will just have all these neuropsychiatric disorders listed with the probability of it. What are we going to do? So, of course, there will be a lot of consequences of these powerful tools being available, but I'm not quite sure about the solution because language by design is something to share. Why do we talk? To share information, so we cannot suppress it.

It's very different from DNA test that you can just say, "No, I'm not going to share it." But language, the definition is just to communicate with people. So, I do not have answers. I can definitely talk to you about all of the potential ethical consequences of it, but I'm not sure I have solutions. So, there should be conversations just focusing on this question.

Lauren Gerlach:

And in terms of this line of research, what's next for you and your team?

Neguine Rezaii:

So, I think already-

Lauren Gerlach:

Shared a few? Yeah.

Neguine Rezaii:

Exactly. You asked the right questions. So, what I'm focusing on is older patients, other languages. So, these are the main topics, and also predicting the future from baseline language. So, these are the active projects that I'm working on. And of course, people now are interested, now that Alzheimer's disease was so promising, let's try other disorders that we... For example, seizure is another disorder that one of my wonderful colleagues approached me, Alice Lam, and we're working on seizure. Can we predict seizure? Can we prognosticate seizure disorders? So, extending it to other languages and other conditions and older ages.

Matt Davis:

We'll wrap this up with Lauren recalling the Craft Story. Neguine, thanks so much for joining us.

Neguine Rezaii:

Lovely. It was a real pleasure. Thank you so much.

Matt Davis:

If you enjoyed our discussion today, please consider subscribing to our podcast. Other episodes can be found on Apple Podcasts, Spotify, and SoundCloud, as well as directly from us at capra.med.umich.edu, where a full transcript of this episode is also available. On our website, you'll also find links to other resources we've created specifically for dementia research. Music and engineering for this podcast was provided by Dan Langa. More information is available at www.danlanga.com. Minding Memory is part of the Michigan Medicine Podcast Network. Find more shows at michiganmedicine.org/podcasts.

Support for this podcast comes from the National Institute on Aging at the National Institutes of Health, as well as the Institute for Healthcare Policy and Innovation at the University of Michigan. The views expressed in this podcast do not necessarily represent the views of the NIH or the University of Michigan. Thanks for joining us, and we'll be back soon.