summary: By analyzing voice, it is possible to predict with 78.5% accuracy whether people with mild cognitive impairment will develop Alzheimer's disease. The tool could make it easier to diagnose and screen for cognitive impairment earlier without the need for expensive tests.
The model uses machine learning to evaluate audio content and can monitor dementia risk in a non-invasive way. Further research aims to improve and extend this technology.
Key Facts:
- High accuracyAI model predicts Alzheimer's disease progression with 78.5% accuracy.
- Accessible Screening: Diagnosing dementia could become easier and more widely available.
- Voice Analysis: Analyzes audio content, not just acoustic features.
sauce: Boston University
Testing for Alzheimer's typically involves a battery of tests, including interviews, brain imaging, blood tests and cerebrospinal fluid analysis, but by that point it may already be too late: memories have already started to fade and long-held personality traits have begun to subtly change.
Early detection can slow the progression of the disease with new breakthrough treatments, but there is no way to reliably predict who will develop dementia associated with Alzheimer's disease.
Now, researchers at Boston University say they've designed a promising new artificial intelligence computer program, or model, that could help change that in the future, simply by analyzing patients' speech.
Their model can predict with 78.5% accuracy whether someone with mild cognitive impairment is likely to remain stable over the next six years or to progress to Alzheimer's disease-related dementia.
As well as helping clinicians predict and diagnose earlier, the researchers say automating parts of the process could also help make screening for cognitive impairment more accessible, without the need for expensive lab tests, imaging or even a doctor's visit.
The model is made possible by machine learning, a subset of AI in which computer scientists teach programs to analyze data on their own.
“We wanted to predict what's going to happen over the next six years, and we found that we could reasonably make that prediction with a relatively high degree of confidence and precision,” says Ioannis (Yiannis) Pashalidis, director of Boston University's Rafik B. Hariri Institute for Computing and Computational Science and Engineering.
“It shows the power of AI.”
A multidisciplinary team of engineers, neurobiologists, computer scientists and data scientists have Alzheimer's and DementiaJournal of the Alzheimer's Association.
“Like everyone, we hope to see more and more treatments for Alzheimer's,” said Paschalidis, a distinguished professor in the Boston University School of Engineering and a founding member of the Department of Computing and Data Science.
“If we can predict what's going to happen, it gives us the opportunity and the time to intervene with drugs to at least keep the disease stable and prevent it from progressing to more severe dementia.”
Calculate the odds of Alzheimer's
To train and build their new model, the researchers used data from the Boston University-led Framingham Heart Study, one of the oldest and longest-running studies in the United States.
Although the Framingham Study focuses on cardiovascular health, participants who show signs of cognitive decline underwent regular neuropsychological testing and interviews, providing a wealth of longitudinal information about their cognitive health.
Pachalidis and his colleagues obtained audio recordings of initial interviews with 166 people, ages 63 to 97, who had been diagnosed with mild cognitive impairment. Of these, 76 were predicted to remain stable over the next six years, while 90 were predicted to experience gradual decline in cognitive function.
They then combined voice recognition tools, similar to the programs that power smart speakers, with machine learning to train a model to find connections between voice, demographics, diagnoses and disease progression. After training the model on a subset of the study population, they tested its predictive ability on the remaining participants.
“We combine the information extracted from the voice recordings with some very basic demographic information like age, gender, etc. to come up with a final score,” Paskalidis says, “which you can think of as the likelihood, the probability, that someone will remain stable or progress to dementia. It has quite a bit of predictive power.”
Rather than using acoustic features of speech such as pronunciation or rate, the model extracts only the content of the interview: the spoken words and their organization.
Pashalidis also said that the information fed into the machine learning program was crude: the recordings were messy, low-quality, and full of background noise, for example.
“These are very casual recordings,” he says, “and yet the model can make something out of this mess.”
This is important because the project was aimed in part at testing AI's ability to make the dementia diagnosis process more efficient and automated, with little to no human involvement.
The researchers say that in the future, models like theirs could be used to deliver care to patients who aren't near a medical center and provide regular monitoring through interactions with an app at home, significantly increasing the number of people getting tested.
According to Alzheimer's International, the majority of people with dementia around the world are not formally diagnosed and go without treatment or care.
Rhoda Oh, a co-author of the paper, said AI has the power to create “equal opportunity in science and medicine.” The study builds on earlier work from the same team, which found that AI can accurately detect cognitive impairment using voice recordings.
“Technology can overcome the stigma of tasks that only those with resources can perform, and care that relies on expertise that not everyone has access to,” said Oh, a professor of anatomy and neurobiology at Boston University's Chobanian Avedisian School of Medicine.
For her, one of the most interesting findings was that “cognitive assessment methods that have the potential to be maximally inclusive, regardless of age, gender, education, language, culture, income and geography, could serve as potential screening tools to detect and monitor symptoms associated with Alzheimer's disease.”
Diagnosing dementia at home
In future studies, Paskalidis hopes to use data not only from formal doctor-patient interviews, with routine questions and predictable interactions, but also from more natural, everyday conversations.
He's already looking at a project to see whether AI can help diagnose dementia via a smartphone app, and he also plans to expand his current research beyond voice analysis (the Framingham test also includes data on patients' drawings and daily life patterns) to improve the model's predictions.
“Digital is the new blood,” Oh says. “We can collect it, analyze what we know today, store it, and reanalyze it tomorrow if something new comes along.”
Funding: The research was funded in part by the National Science Foundation, the National Institutes of Health, and the Rajen Kilachand Fund for Integrative Life Sciences and Engineering at Boston University.
About this AI and Alzheimer's research news
author: Katherine Gianni
sauce: Boston University
contact: Katherine Gianni – Boston University
image: Image courtesy of Neuroscience News
Original Research: Open access.
“Predicting Alzheimer's disease progression within 6 years using speech: A new approach using language models” by Ioannis Paschalidis et al. Alzheimer's and Dementia
Abstract
Predicting Alzheimer's disease progression within 6 years using speech: A new approach using language models
introduction
Identifying individuals with mild cognitive impairment (MCI) who are at risk of developing Alzheimer's disease (AD) is essential for early intervention and clinical trial selection.
Method
We applied natural language processing and machine learning techniques to develop a method to automatically predict progression to Alzheimer's disease within six years using voice samples. The study design was assessed with a neuropsychological testing interview. yeah = 166 participants from the Framingham Heart Study (90 with progressive MCI and 76 with stable MCI)
result
Our best model, using features generated from speech data, age, sex, and education level, achieved 78.5% accuracy and 81.1% sensitivity for predicting progression from MCI to AD within 6 years.
Discussion
The proposed method provides a fully automated procedure and offers the opportunity to develop an inexpensive, widely accessible and easy-to-administer screening tool for predicting progression from MCI to AD, facilitating the development of remote assessments.
