Dose of learning: AI will be able to more accurately detect Alzheimer's disease in the early stages

Artificial intelligence will be able to detect signs of Alzheimer's disease even with a lack of genetic data. Scientists from Harvard have already applied the Russian development: the use of artificially created data has helped to increase the accuracy of recognition of cognitive impairments by up to 77%. Technology can expand the possibilities of medical research, in which each data set plays an important role. According to doctors, artificial intelligence can significantly change approaches to the diagnosis of neurodegenerative diseases, shifting the focus from the detection of irreversible disorders to early prevention. However, so far the algorithm has recognized cognitive impairments rather than confirming the diagnosis of Alzheimer's disease. Therefore, it can only be used as a tool to support medical decision-making, neurologists believe.

AI training with a lack of genetic data

Scientists from Harvard University and the University of Texas have found a way to improve the work of artificial intelligence in the study of Alzheimer's disease, using technology from Yandex and the HSE Faculty of Computer Science. One of the main difficulties of such studies is the lack of genetic data from patients needed to train neural networks, the developers explained. Researchers have shown that this problem can be partially solved with the help of artificially created data. This approach has improved the accuracy of recognizing cognitive impairments associated with the disease.

Photo: Global Look Press/IMAGO/Iuliia Zavalish

For the experiment, the scientists used the genetic and demographic data of 542 people. Neural networks are able to identify complex relationships between body features and diseases, but they usually require a large number of examples to teach them. If there is not enough data, the algorithm risks remembering the characteristics of specific patients instead of learning to recognize the signs of the disease in other people.

The open TabDDPM technology helped solve this problem. This is a method of generating artificial data using a neural network that studies patterns in real information and creates new statistically similar records based on them, rather than simply copying existing ones. In this case, scientists have obtained additional combinations of genetic and demographic characteristics. This allowed us to increase the amount of data for AI training without attracting new research participants.

—We have published the TabDDPM source code in the public domain, so researchers and developers from any country can use the technology to train their own models," Artem Babenko, PhD in Physics and Mathematics, head of the Yandex Research laboratory and head of the Yandex Scientific and Educational Laboratory at the Higher School of Economics, told Izvestia.

Photo: IZVESTIA/Dmitry Korotaev

The researchers compared the work of an AI trained only on real data with a model that additionally used artificially created information. In the second case, the neural network distinguished people with and without cognitive impairments more accurately: the recognition quality index (AUC) increased from 0.721 to 0.769, and the proportion of correctly identified cases of impairments increased from 68 to 77%. The results showed that synthetic data can improve the effectiveness of medical AI training, but they are not a substitute for observing real patients.

The path to the unconscious: how not to miss the signs of dementia in a loved one

Doctors have named alarm signals that are worth paying attention to.

Application of new technology in medicine

The algorithm does not make diagnoses on its own, but helps to train neural networks that can later be used for diagnosis, the developers explained. Thanks to artificially created data, it is possible to increase the training sample without involving additional patients. This makes it possible to improve the quality of models and potentially reduce the likelihood of errors in disease detection.

Artificial intelligence can radically change approaches to the diagnosis of Alzheimer's disease, shifting the focus from the detection of already irreversible disorders to early prevention, Timofey Voronin, deputy director of the Research Center for Strategic Analysis of Technological Development at the Higher School of Economics of Lomonosov Moscow State University, told Izvestia.

Photo: IZVESTIA/Sergey Lantyukhov

— Unlike traditional tests, machine learning algorithms analyze arrays of data, including MRI, PET, and genetic biomarkers, objectively identifying pathologies several years before the onset of clinical symptoms. For patients, this means a quick and non—invasive procedure, and for doctors, it means eliminating the subjective factor and obtaining a critically needed time window to start therapy," he noted.

The expert also stressed that the demand for such solutions will only grow in the global economy. The global shortage of qualified radiologists is becoming one of the drivers of the development of the medical AI market, which already exceeds $2 billion in 2026.

— I believe that in the coming years, the routine decoding of images and the preparation of basic medical reports will be fully transferred to neural networks. This will allow the healthcare system to redirect human resources to analyzing complex clinical cases, verifying data, conducting research, operations and other medical manipulations," said Timofey Voronin.

The main value of the proposed technology lies in the ability to improve the accuracy of artificial intelligence algorithms when working with limited medical data, said Lusine Brsikyan, a neurologist at the 5th Neurological Department of the Russian Center for Neurology and Neuroscience, an expert at the Brain Health and Development Foundation.

Photo: IZVESTIA/Yulia Mayorova

— Alzheimer's disease is a complex neurodegenerative disease, the diagnosis of which requires a comprehensive assessment of cognitive functions, neurological status, neuroimaging results and, if necessary, specific biomarkers. The use of synthetic data allows neural networks to be trained on more diverse combinations of genetic and demographic traits, even if the initial sample of patients is relatively small. In the future, this may contribute to the development of more accurate tools for detecting cognitive impairments, including in the early stages," the doctor noted.

Signal with help: the "commander" molecule may be the key to restoring the brain

Researchers have found a new target for the treatment of Alzheimer's and Parkinson's diseases

It is important to keep in mind that in this study, the algorithm recognized cognitive impairments rather than confirming the diagnosis of Alzheimer's disease. Therefore, it is premature to talk about creating a ready-made diagnostic method, the neurologist explained. In the future, such algorithms may become an additional tool to support medical decision-making. For example, they will help identify patients with an increased likelihood of cognitive impairment who need a more in-depth examination, as well as take into account the complex relationships between genetic risk factors and clinical manifestations of the disease.