A close-up shot of a microphoneShare on Pinterest
Researchers are using AI to analyze the speech patterns of people with Alzheimer’s. MirageC/Getty Images
  • Researchers at Boston University say they have designed an artificial intelligence tool that can predict with nearly 80% accuracy whether someone is at risk for developing Alzheimer’s disease based on their speech patterns.
  • The ability to identify potential cognitive decline early has significant potential for mitigating the progression of Alzheimer’s, experts say.
  • However, the sample size used was small, and experts caution that such a tool is not meant to be leaned on as an exclusive method.

Researchers at Boston University say they have designed an artificial intelligence tool that can predict with nearly 80% accuracy whether someone is at risk for developing Alzheimer’s disease based on their speech patterns.

Using a natural language processing model, the researchers attempted to see if people with mild cognitive decline would develop Alzheimer’s within a six-year period. They focused on a cohort of 166 people — 107 women and 59 men — between the ages of 63 and 97 who had some level of cognitive complaints.

Each participant had been part of the Framingham Heart Study, which is led by Boston University, and as part of the study had been recorded for an hourlong interview. Those interviews formed the data that was analyzed by the AI tool the researchers developed.

Of the cohort, 90 people went on to have progressive declines in their cognitive function, while 76 remained stable. But what the Boston University researchers discovered was that by combining speech-recognition tools and machine learning, they could track connections between speech patterns and cognitive decline, based on biomarkers associated with cognitive decline. The model they developed, albeit in a small sample size, was able to predict significant cognitive decline with 78.5% accuracy, the researchers say.

Melissa Lee, PhD, the assistant director of the Diagnostics Accelerator at the Alzheimer’s Drug Discovery Foundation, told Medical News Today that the accuracy of the researchers’ findings was a very positive sign for the future.

“The fact that they’re showing such high accuracy off of such a small data set is actually really promising and shows that if they were to use a much larger data set, like SpeechDx, one of the datasets the ADDF has been developing, then I think there is much more potential for growth here. Of course, it’s hard to say what the ceiling is there — it’s hard to put kind of an upper limit on how accurate it can be,” Lee said.

“With a larger, more thorough characterized dataset, it could probably get quite accurate,” she added.

Dementia directly affects more than 55 million people worldwide, and up to 70% of those people have Alzheimer’s disease, which is characterized by a loss of brain cells associated with the toxic buildup of two proteins, amyloid and tau.

The most common symptoms of Alzheimer’s disease are memory loss, cognitive deficits, problems with speaking, recognition, spatial awareness, reading, or writing, and significant changes in personality and behavior. Since Alzheimer’s is progressive, these symptoms are usually mild at first and tend to become more severe over time. With no cure for the disease, patients and caregivers must approach treatment with medication, lifestyle changes, and support groups.

Emer MacSweeney, MD, CEO and consultant neuroradiologist at Re:Cognition Health in London, England, told Medical News Today that using artificial intelligence to detect Alzheimer’s early offers doctors and patients potentially more options than currently at their disposal.

“The AI tool for predicting Alzheimer’s progression offers several significant advances: it enables early intervention with treatments to slow the disease, improves accessibility to cognitive assessments through automated and remote screening, and facilitates personalized care plans based on predicted disease trajectories,” MacSweeney said.

“Additionally, it helps healthcare providers prioritize patients needing intensive monitoring, optimizing resource allocation, and provides valuable data for refining predictive models and developing new treatment strategies,” he added.

Lee said that the current wave of drugs for Alzheimer’s work best when prescribed as early as possible, but she added that detection of potential cognitive decline is often based on the presence of amyloid in the brain. But she said that a large portion of people who test positive for amyloid in the brain don’t go on to develop cognitive symptoms — so the AI model can provide more accurate predictions.

“Having the ability to say in the future, you’re likely to decline, or you’re not likely to decline, gives you the ability to intervene much more early than we possibly can, either through the drugs that are currently being developed today, that target things like amyloid, or via the things that we already know to be impacting risk for developing Alzheimer’s later on,” Lee said.

“We know that 40% of Alzheimer’s cases today can be delayed or prevented through basic lifestyle modifications: things like monitoring, eating a heart-healthy diet, reducing alcohol consumption, things like that. Exercise, treating depression — these are things that if you know down the line someone has an increased risk for developing Alzheimer’s disease you can go ahead and say ‘these are things you can be doing to manage that risk.’”
— Melissa Lee, PhD

MacSweeney pointed out that even though a nearly 80% accuracy rate is high, there is still the potential for false positives or negatives, and automating healthcare in such a way could create problems.

“There will be cases where the tool incorrectly predicts disease progression or stability, potentially leading to undue stress or false reassurance,” MacSweeney said.

“While the use of AI has the capability to revolutionize certain areas of healthcare, there is a risk that clinicians might over-rely on AI predictions without considering the broader clinical context, leading to potential misdiagnoses,” she cautioned.

Lee also cautioned that while the accuracy rate is promising, the sample size was small, and ultimately the tool is not meant to be used on its own — now or in the future.

“It might, in cases when it might be wrong, it might raise a lot of alarm, cause a lot of stress to an individual unnecessarily. But I think this tool is very likely not meant to be used alone, but meant to be used in conjunction with a whole bunch of other things like a blood test that can give you a more accurate picture or holistic image of how an individual is doing that would then kind of mitigate some of that risk,” she said.

“Speech is something that’s really easy to gather from a person, and you can gather that really easily over time. Like right now, if someone is recording this phone call, they have a sample of my speech. If they record another phone call I do next week, they have a sample of my speech over time. What that means is that you can look at how an individual’s speech changes from one moment to the next. And that gives you kind of an individual baseline so you understand what is normal for a person and what is abnormal for that person specifically.”
— Melissa Lee, PhD