Artificial intelligence for early detection of Alzheimer’s disease

In 1950, Alan Turing, considered by some to be the father of artificial intelligence, wondered how to make a computer intelligent. Those pioneers of the 50s diagrammed in a theoretical way what is applied everywhere today: using networks of neurons as a model so that a computer can be intelligent. That is, emulate the interconnection of neurons in the human brain.

More than seventy years later, researchers from the National University of Buenos Aires (UBA) in Argentina use these artificial neural networks so that artificial intelligence learns to identify the onset of Alzheimer’s disease in magnetic resonance images. At the same time, seeing these artificial neural networks in action allows unparalleled freedom to understand how the human brain works.

“One advantage of working with an artificial neural network is that we can take it apart and dissect it however we want to study it down to the smallest detail. Those of us who do artificial intelligence from neuroscience seek that the artificial allows us to learn about the structure of the human brain, and vice versa. How things are codified, it’s a round trip,” explained Diego Fernandez Slezak, a researcher at the Institute of Computer Sciences (UBA / National Council for Scientific and Technical Research (CONICET) in Argentina) and professor at the Faculty of Exact Sciences. and Natural Sciences from the University of Buenos Aires.

Fernandez Slezak was recently awarded the 2023 Konex Science and Technology Award. He also received the Google Research Award in 2016, both for the work he has been doing with artificial intelligence applied to neuroscience for more than ten years.

“Now we are working with images and how to code, with magnetic resonances of the brain, attributes that reveal the imminence of Alzheimer’s disease,” said the expert. “It is a very well characterized disease, and it is under constant study. Currently, there are very good treatments to delay the effects that affect quality of life, so detecting it early is vital”.

The work carried out by Fernández Slezak and colleagues from the Institute of Computing Sciences of the Faculty of Exact and Natural Sciences of the UBA is to identify a pattern in the magnetic resonance images of the brain that allow artificial intelligence to identify the disease early Alzheimer’s.

“We use artificial intelligence to code the brain,” explained the researcher. “When an MRI is done on a person’s brain, what we see are colored pixels or voxels, which are little pieces of the brain that light up or don’t light up depending on what is happening in that area. If what lights up is normal, it is difficult to identify, even for specialists. Then, artificial intelligence takes the images, encodes them, and returns indices that could allow them to be associated with pathologies, such as Alzheimer’s, bipolarity, stroke, or epilepsy.”

This is an ongoing investigation in collaboration with the Fleni Hospital, which is the one that provides the magnetic resonances with which artificial intelligence is being taught. In order for you to learn the whole path that leads to a disease like Alzheimer’s, it begins with what is known as a patient’s subjective complaint. That is to say, someone who forgets or confuses names.

The specialists then carry out a battery of studies on the patient, including magnetic resonance imaging, to find out if this subjective complaint actually corresponds to a real brain problem. This could be the beginning of a dementia that could later degenerate into Alzheimer’s disease, or perhaps aphasia, or another brain pathology.

The researchers use, then, all the resonances that were made to a person who later detected something, to train artificial intelligence to identify if there is a previous pattern that allows detecting those diseases in time.

“El Fleni has a database of thousands of old MRIs,” the expert told Argentina Investiga. “What we do first is that artificial intelligence encodes the structure of the brain, and then what we have to do is associate that structure with the different pathologies. You look at the diagnostic history, and you see if a given direction of brain structure is the direction that leads to Alzheimer’s disease, or to each of the other pathologies.”

Images of the brain captured by techniques such as MRI and others can be analyzed with artificial intelligence to find signs of impending neurological disease. (Image: NIAAA)

Artificial neurons to understand the real ones

Artificial intelligence works with what is known as neural networks, including ChatGPT, so famous today for answering all kinds of queries. It is a method that allows artificial intelligence to process data in a way inspired by how the human brain does.

It uses nodes that are interconnected in a layered structure, much like how biological neurons work. It is a system that allows computers to learn from their mistakes and continually improve. They establish relationships between the data with which they are fed, providing complex results.

“The way these artificial neural networks learn has to do with how we represent the information in our heads,” Fernandez Slezak said. “Our idea is to study how it manages to do that in very small niches and where we have some control. Language is one of those areas, that is, how does an artificial intelligence learn to write?

“The free association between words, which is something automatic in us, also occurs automatically in these artificial networks. This allows us to see how ideas are related in the brain”, said the expert. “We are doing that with the artificial neural network to see what parallels there are with the functioning of the human mind.”

“While we can give people MRIs to see how their brain works when it associates words, we can’t get the level of detail that allows us to study the neural network when it does the same. Then it can help us understand how information is represented in our brain”, explained Fernández Slezak.

The interesting thing about the parallelism between how artificial intelligence processes information and the human brain is that there is precisely a lack of knowledge about how it works. It is not known why a hundred thousand balls or nodes connected with weights, which are probabilities of connection between one and the other, manage to represent the language well, like an artificial intelligence in the style of ChatGPT. Nor is it clear how billions of neurons make a human being represent language.

It is known that the mathematical algorithm of an artificial neural network adapts the weights of the nodes or neurons, but it is unknown why a network structure with certain weights solves a given problem very well. The same for its biological parallel. That is why it is useful to be able to understand the way in which an artificial intelligence “thinks”, to finish understanding how and why the human being thinks as he thinks. (Source: Martín Cagliani / UBA / Argentina Investiga)