A machine learning model that can detect signs of Alzheimer’s across language

Researchers are working to enable early diagnosis of Alzheimer’s disease using machine learning (ML) models. This model could one day turn into a simple screening tool for anyone with a smartphone.

The model can distinguish Alzheimer’s patients from healthy controls with an accuracy of 70-75 percent, a promising figure for the more than 747,000 Canadians with Alzheimer’s disease or other forms of dementia. became.

Dementia of the Alzheimer’s type can be difficult to detect in its early stages because it often begins with very subtle symptoms that can be confused with memory-related problems that are typical of older people. . But, as the researchers point out, the earlier potential problems are detected, the sooner patients can take action.

“Previously, detecting changes in the brain required lab work and medical imaging, which was time consuming, expensive, and no one had been tested so early.” Eleni Straulia, a professor of computer science who was involved in the study. when creating the model.

“If we can use mobile phones to get early signs, we can get to know the patient-doctor relationship, which can potentially help start treatment sooner, and we can use mobile devices at home. We can also initiate simple interventions to slow the progression.”

Screening tools do not replace medical professionals. However, in addition to helping with early detection, telemedicine is a convenient way to identify potential concerns for patients who may face geographic or language barriers in accessing services within their community. Born, explains Zela Shah, a master’s student in the Department of Computing. Scientist and lead author of the paper.

“With this kind of technology, you can think about triaging patients based entirely on their voice,” says Shah.

While the research group has previously investigated the language used by people with Alzheimer’s disease, in this project they investigated language-independent acoustic and verbal speech characteristics rather than specific words.

“The original job involved listening to people and making sense of what they said, which is an easier computational problem to solve,” says Stroulia. “Now we are saying listen to the voice. There are some characteristics of the way people speak that go beyond language.”

“This is much stronger than the version issues we were solving before,” Stroulia adds.

The researchers started with a language trait that doctors noted was common in people with Alzheimer’s disease. These patients tended to speak more slowly and had more pauses and breaks in speech. They usually used shorter words and often had less clarity of speech. Researchers have found a way to transform these features into speech features that the model can screen.

The researchers focused on English and Greek speakers, but “the technology could be used in a variety of languages,” Shah said.

The model itself is complex, but the end user experience of the tool that incorporates it is even simpler.

“When a person talks to the tool, it analyzes it and makes a prediction. Either the person has Alzheimer’s disease or they don’t,” said the paper’s contributor and computer science professor. One Russ Greiner says: . That information is provided to health care professionals so they can determine the best course of action for the person.

Greiner and Stroria both lead the Computational Psychiatry Research Group at the University of Australia, whose members have developed similar AI models for detecting mental illnesses such as PTSD, schizophrenia, depression and bipolar disorder. and tools.

“Anything we can do to enhance the clinical process, deliver treatments faster at less cost, and manage the disease is great,” says Stroria.

The ML model is described in the paper “Exploring Language-Independent Speech Representations Using Domain Knowledge for Detection of Alzheimer’s Disease”. ICASSP 2023 Signal Processing Grand Challengethe team was ranked #1 in North America and #4 in the world.