Automated prediction of Alzheimer’s disease progression using speech and machine learning

Trying to figure out if someone has Alzheimer’s usually involves a series of evaluations—interviews, brain imaging, blood and cerebrospinal fluid tests. But, by then, it’s probably already too late: memories have begun to slip away, long-established personality traits have begun to subtly change. If caught early, new breakthrough treatments can slow the relentless progression of the disease, but there’s no sure way to predict who will develop Alzheimer’s-related dementia.

Now, Boston University researchers say they’ve designed a promising new artificial intelligence computer program, or model, that could one day help change that—simply by analyzing a patient’s speech.

Their model can predict, with an accuracy rate of 78.5 percent, whether someone with mild cognitive impairment is likely to remain stable over the next six years — or develop Alzheimer’s disease-related dementia. While allowing clinicians to look into the future and make earlier diagnoses, the researchers say their work could also help make screening for cognitive impairment more accessible by automating parts of the process — no expensive lab tests, imaging tests or even office visits. The model is powered by machine learning, a subset of artificial intelligence where computer scientists teach a program to independently analyze data.

We wanted to predict what would happen over the next six years – and found that we can reasonably make that prediction with reasonably good confidence and accuracy. It shows the power of artificial intelligence.”

Ioannis (Yiannis) Paschalidis, Director BU Rafik B. Hariri Institute for Computing and Computational Science & Engineering

The interdisciplinary team of engineers, neurobiologists, and computer and data scientists published their findings in Alzheimer’s & Dementiathe journal of the Alzheimer’s Association.

“We hope, as everyone does, that more and more treatments for Alzheimer’s will become available,” says Paschalidis, a distinguished professor of engineering in the BU College of Engineering and a founding member of the School of Computer & Data Science. “If you can predict what’s going to happen, you have more opportunity and time to intervene with drugs and at least try to keep the condition stable and prevent progression to more severe forms of dementia.”

Calculating the probability of Alzheimer’s disease

To train and build their new model, the researchers turned to data from one of the nation’s oldest and longest-running studies — the BU-led Framingham Heart Study. Although the Framingham Study focuses on cardiovascular health, participants who show signs of cognitive decline undergo regular neuropsychological testing and interviews, producing a wealth of longitudinal information about their cognitive well-being.

Paschalidis and his colleagues recorded 166 initial interviews with people between the ages of 63 and 97 who had been diagnosed with mild cognitive impairment—76 who would remain stable over the next six years and 90 whose cognitive function would gradually decline. They then used a combination of speech recognition tools–similar to the programs that power the smart speaker–and machine learning to train a model to find connections between speech, demographics, diagnosis, and disease progression. After training it on a subset of the study population, they tested its predictive ability on the rest of the participants.

“We combine the information we get from the recordings with some very basic demographics—age, gender, etc.—and get the final score,” Paschalidis says. “You can think of the score as the probability, the probability, that someone will remain stable or move into dementia. It had significant predictive power.”

Rather than using auditory features of speech, such as diction or speed, the model simply draws from the content of the interview—the words that are spoken, how they are structured. And Paschalides says the information they put into the machine learning program is raw: recordings, for example, are messy, low-quality and full of background noise. “It’s a very casual recording,” he says. “And yet, with this dirty data, the model can make something out of it.”

This is important because the project was in part about testing the ability of AI to make the dementia diagnosis process more efficient and automated, with little human involvement. In the future, the researchers say, models like theirs could be used to provide care to patients who are not located near medical centers or to provide routine monitoring through interaction with an app at home, drastically increasing the number of people who are subject to a preventive check. According to Alzheimer’s Disease International, the majority of people with dementia worldwide never receive a formal diagnosis, thereby being excluded from treatment and care.

Rhoda Au, a co-author on the paper, says AI has the power to create “equal opportunity science and healthcare.” The study builds on previous work by the same team, where they found that AI could accurately detect cognitive decline using voice recordings.

“Technology can overcome the bias of work that can only be done by those with resources or care that relies on specialized expertise that is not available to everyone,” says Au, a professor of anatomy and neurobiology at the BU Chobanian & Avedisian School of Medicine. . For her, one of the most exciting findings was “that a method of cognitive assessment that has the potential to be as inclusive as possible—?perhaps independent of age, sex/gender, education, language, culture, income, geography—could serve as a potential screening tool to identify and monitor symptoms associated with Alzheimer’s disease.”

A dementia diagnosis from home

In future research, Paschalides would like to investigate using data not only from formal clinician-patient interviews—with their scripted and predictable back-and-forth questions—but also from more natural, everyday conversations. He is already looking at a project on whether artificial intelligence can help diagnose dementia through a smartphone app, as well as expanding the current study beyond speech analysis – the Framingham trials also include patient designs and data on patterns of daily life – to enhance the model’s predictive accuracy.

“Digital is the new blood,” says Au. “You can collect it, analyze it for what’s known today, store it, and analyze it again for anything new that comes up tomorrow.”

This research was funded, in part, by the National Science Foundation, the National Institutes of Health, and the BU Rajen Kilachand Fund for Integrated Life Science and Engineering.