OpenAI’s generative artificial intelligence could serve as the basis for new tools to fight this terrible disease.
For the past few weeks, the web has only had eyes for ChatGPT. The incredible machine learning-powered chatbot designed by OpenAI has already shown that it can hold conversations, write fictional texts, help students cheat… and one day it may be able to participate in the fight against Alzheimer’s disease.
It is a neurodegenerative disease with terrible consequences for patients and their families. It typically results in loss of memory and lucidity that can take on catastrophic proportions over time. Patients then tend to express themselves in a disjointed, incoherent or even barely intelligible way.
The problem is that despite promising leads, there is still no treatment as such. Admittedly, there are many protocols that make it possible to combat the symptoms, but we are still unable to completely cure this scourge.
The challenge is therefore to arrive at this diagnosis, which has serious consequences, as soon as possible. Indeed, the effectiveness of the treatment depends a lot on the moment of treatment. If the disease is identified in the early phase, it is possible to considerably slow down its progression.
But then we are faced with another problem. Because even when you are particularly vigilant, it is not always easy to identify Alzheimer’s disease in the early phase. Indeed, the symptoms tend to appear relatively late. ” Studies have shown that by the time the diagnosis is made, up to 90% of brain cells may already be dead “explained Valerie Daggett, professor of bioengineering at the University of Washington who developed a blood test to diagnose the disease.
Statistical tools to track the disease
Blood test detects Alzheimer’s years before symptoms
To circumvent this problem, the most obvious solution would be to set up a systematic follow-up of the population, in particular in patients at risk. Because to give yourself every chance of reaching the right diagnosis, you typically have to go through brain imaging sessions and extensive cognitive assessments. It would therefore be an expensive, time-consuming monitoring protocol, and therefore difficult to apply on a large scale.
A team of researchers from the University of Drexel, in the United States, opted for another approach in order to carry out this preventive work. Instead of looking for physiological markers of the disease, such as the famous amyloid plaques, they plan to go through statistical analysis of language.
The idea is to dissect the person’s speech using computer tools to identify extremely discreet warning signs. Generally, this involves looking for a few strictly acoustic cues; in other words, we are interested in the way the individual expresses himself rather than in the deeper meaning of his remarks. This approach works quite well. In about 75% of casesthis makes it possible to reach a diagnosis before the onset of symptoms, which considerably improves management.
However, it turns out that artificial neural networks are particularly good at highlighting these discrete statistical deviations. And in parallel, in recent years, AI tools applied to language have been gaining momentum at a staggering speed. This offers a new possibility to researchers: instead of focusing only on the acoustic dimension, it is now possible to look directly at the meaning of the words.
Generative AI to the rescue
The Drexel researchers therefore asked themselves a timely question: could a program like ChatGPT detect the first signs of Alzheimer’s disease, long before human specialists?
“These language models like GPT3 are so powerful that they can spot these kinds of subtle differences,” says Hualou Liang, lead author of the study. “If the subject has Alzheimer’s disease and it is already felt in the language, we hope to be able to use machine learning to offer an early diagnosis”, he specifies.
To test this hypothesis, they started by collecting 237 audio recordings. They came from two groups of patients, either healthy or suffering from Alzheimer’s disease. According to IEEE Spectrum, to analyze them, they relied on an unknown feature of ChatGPT.
Indeed, the program is not only able to offer a textual response. It can also produce an “embedding”. In essence, it’s a very rich mathematical representation of a piece of text. Normally, this allows the algorithm to compare the meaning of two snippets. Here, the researchers have diverted it to analyze the coherence of the comments.
And the results have been quite remarkable. In 80.3% of cases, based solely on embedding, the team succeeded in determining whether or not the targeted patient belonged to the Alzheimer’s group. A not-yet-revolutionary, but quite significant increase over what traditional acoustic analytical methods offer.
Towards new diagnostic tools based on AI
There are two conclusions to be drawn from this study. The first is that it confirms the potential of AI applied to language in the management of Alzheimer’s. This is excellent news, because the researchers have thus paved the way for the development of new diagnostic tools that are not only very powerful, but also much less cumbersome than the usual protocols.
The second conclusion is that we will probably have to be very careful before having developed specialized solutions of this kind. Because it should also be remembered that ChatGPT is commercial software developed by a private company. And according to Frank Rudzicz, a specialist also interviewed by IEEE Spectrum, this poses a big problem of transparency and accessibility.
” These closed APIs are limited, because you can’t inspect the code or make deep changes to it. “, he explains. “ This means that we cannot carry out series of more advanced experiments which make it possible to find the source of potential errors which will necessarily have to be corrected or avoided. “, he regrets.
Liang is also quite transparent about the limits of this work. He insists on the fact that for all these reasons, it is not yet a tool adapted to clinical issues; it is not tomorrow the day before that specialists in neurodegenerative diseases will hand over to ChatGPT.
But the most important thing, and what should be learned from this work, is the potential of this approach. Once specialized tools emerge and mature, humanity will have access to a tremendously powerful tool to wage an unprecedented war against Alzheimer’s. Knowing the human and medical stakes of this scourge of public health, this is progress that should be applauded without restraint. And that’s probably just the beginning, knowing the incredible potential of artificial intelligence in many branches of medicine.
The text of the study is available here.
