New linguistic tech is able to detect Alzheimer’s
July 12, 2023
EDMONTON – Computer science researchers at the University of Alberta are branching into linguistics and healthcare to develop accessible detection technology for Alzheimer’s dementia. They’ve created a machine learning (ML) model that has detected Alzheimer’s with 70 to 75 percent accuracy. The ML model could provide easier access to Alzheimer’s screening and allow for earlier treatment, according to Zehra Shah (pictured).
Shah is a master’s student in the department of computing science. She is also the first author on the paper that describes the ML model. With a team of researchers, Shah developed the ML model for the International Conference on Acoustics, Speech, and Signal Processing (ICASSP) Signal Processing Grand Challenge (SPGC) 2023. The U of A placed first in North America and fourth globally in the challenge.
Shah explained that her team used roughly 200 to 250 speech samples given to them by the SPGC. They used the samples to develop a supervised learning model, which is a type of ML model. The samples were short audio recordings of test subjects describing a picture.
“We had [labeled] speech samples from both dementia patients and healthy controls. That is what constituted our training set [which] we used to teach the computer to look for patterns.”
Once trained, the ML model is able to search for these patterns in unlabelled speech samples. It then predicts whether the sample is a dementia patient or a healthy control.
The challenge required work done in a language agnostic manner. In this manner, models had to process samples effectively in any given language.
“[The] training sample was from English speech of dementia patients and healthy controls, but then the test set was supposed to come from Greek speech samples. What we wanted to do was look for features in their speech which would be transferable across these two languages,” Shah said.
They sorted the spontaneous speech samples into healthy controls and Alzheimer’s dementia patients based on three main features: pauses in speech, speech complexity, and speech intelligibility.
The researchers hypothesized that dementia patients have longer and more frequent pauses in their speech. They also predicted that the patients “would have lower speech complexity, so they are using shorter words more frequently compared to healthy controls.”
Shah said that her team looked at word confidence when analyzing speech intelligibility. “With dementia patients, usually their speech is a little more cultured. It’s harder to understand the words they are speaking.”
The team has previously used this technology for other mental health disorders, including post-traumatic stress disorder (PTSD) and depression.
Shah said that the ML model could identify people at risk of Alzheimer’s dementia earlier on, which would help clinicians and patients. Additionally, it could continuously monitor the dementia’s progression.
The primary benefit is that the ML model makes monitoring Alzheimer’s dementia more affordable and accessible. Shah explained that brain imaging “is used very regularly, but it’s very expensive, it’s resource intensive, [and] it’s not available in many geographic locations.”
With the ML model, “you can imagine a web or smartphone app where we can just screen for whether somebody is at risk of developing dementia or not. Then you can notify them with potential follow-up recommendations and get them to a clinician possibly earlier than what is done right now.”
Though the ML model is “not at the stage where it can be used as a full-blown clinical tool,” Shah sees a lot of potential in it. She emphasized the relevance of her work, since the number of Canadians with dementia is expected to triple by 2050.
“There are so many low resource countries out there that don’t have a lot of mental health care at all. Just this kind of screening, if done at scale, can be pretty beneficial for the global population considering the fact that Alzheimer’s dementia is on the rise globally.”