OsloMet professor Anis Yazidi is working to change that. He’s a member of the OsloMet team leading the AI module of AI-Mind, a program that’s seeking to improve our understanding of the changes in the brain that lead to dementia, and perhaps even bringing us closer to finding a treatment.
From animal mimicry to brains
Yazidi has seen a remarkable improvement in AI throughout his career. He received a master’s in computer science in Norway 2008 and a PhD working alongside some of the world’s leading AI researchers at the Ericsson Research lab in Aachen.
When I started, AI was about adaptive learning, like animals learning simple tricks. Today it’s mimicking the human brain.– Anis Yazidi
The most sophisticated AI today can understand images and videos. Transformers – the main technology behind the popular ChatGPT and similar programs – can find connections between distant parts of an image or between early and later parts of a video.
Researchers can use this ability to identify important correlations in brain scans and other medical data. It’s this ability for AI to learn that Yazidi believes holds the key to understanding dementia.
Predicting health outcomes
For clinicians and patients, dementia is frustratingly inscrutable.
“From a medical point of view, there is no way to detect it in an early phase in order to perform some kind of intervention,” explains Yazidi.
He and the AI-Mind team are using AI to find one.
Their main research focuses on people with mild cognitive impairment (MCI) to determine what factors lead to dementia.
“MCI happens to pretty much everyone eventually,” Yazidi explains.
“You start forgetting names and numbers. Once that happens, there is a certain chance, maybe 30-50 percent, that it will develop into dementia in the next two years.”
The key is figuring out what factors determine whether MCI will progress into dementia or not. That requires a lot of data and the kind of analysis that AI is particularly good at. For the data part, AI-Mind is finding people with MCI and following them for two years, collecting medical information like brain scans, cognitive test results, and blood samples.
Teaching the AI
For the analysis, they use an AI technique called ‘supervised learning’. They tell the algorithm whether patients developed dementia or not and ask it to find ways to distinguish between the groups. Unlike many AI models today, AI-Mind’s model can explain how it made its distinction. This is an important feature if they want to understand the underlying causes. Because the algorithm can process far more data than a human, it may even be able to find connections that the clinicians don’t know about.
A second approach is to use the AI’s ability to measure different physical properties from an electroencephalogram (EEG) such as spectral power, connectivity, and how brain signals correlate. Clinicians from the AI-Mind project have a hypothesis that dementia starts in the electrical signals. If the researchers can study the connectivity and how it breaks down, they may be able to identify the beginnings of dementia so that physicians have a chance to treat it before it’s too late.
Working together
The potential of AI-Mind is possible because it is a massive collaboration. It brings together 15 partners from across four countries, including researchers, clinicians, health economists, product developers, and ethicists.
Yazidi says that managing such a diverse team has been a challenge coming from his computer science background, but the benefits have been invaluable.
If the data is just on your computer, you lose connection to the patients; there are clinical insights that you can get only when you are looking at the at the data together with the clinicians.– Anis Yazidi
These clinicians invited Yazidi and the AI team to come see the patients and get a better understanding of the problem and data collection process. They also lent their experience identifying artifacts in the data so Yazidi didn’t waste valuable time pursuing seemingly promising signals that turned out to just be the result of, as in one example, a particularly hairy participant.
In addition to the clinical expertise, Yazidi works with ethicists who help make sure that their data and algorithm aren’t biased.
“The data needs to be representative or it won’t work,” Yazidi says.
“If you put too many Norwegians in your data, your algorithm will not function on people of different ethnicities; if you include too much from a certain age group, your algorithm will not work for another age group; if you don't have equal representation of sexes, it won’t work for everyone.”
Building tools and solving problems
Yazidi and AI-Mind are planning to present their initial results at a conference in October. For now, they want to get feedback from other AI researchers and clinicians to make sure that they are on the right track.
The next step is developing this algorithm into a dementia screening tool. EEG coupled with their screening software will be an inexpensive and effective first assessment.
“Once you have a good screening tool and you identify a risk group, you can proceed with the more expensive tests” he says. This is especially important for developing countries and places where healthcare access is limited by cost.
While Yazidi cautions that this research will likely take years to bear fruit and the goal of this stage is only prognostic, this research could ultimately show us what causes dementia and give researchers a target for treatment. Until it’s ready to do that, AI-Mind will continue to collect more diverse data and improve its AI algorithms.
