Specialists from the Immanuel Kant Baltic Federal University (IKBFU), in collaboration with the Chinese Northwestern Polytechnical University, have developed a neural network for diagnosing autism. The artificial intelligence determines the disease with a 95% probability. This was reported by the press service of the Russian university.
The new development focuses on the analysis of electroencephalograms (EEGs). To train the neural network, data from 298 children in the age group from 2 to 16 years were used. Most of them suffered from various forms of autism.
The algorithm we developed is aimed specifically at finding distinctive features in the group of children with autism, considering the functional networks of healthy children to be common to both groups. Therefore, it allowed us to find signs that other machine learning algorithms miss. In the future, our proposed machine learning approach will help to identify autism spectrum disorders at earlier stages than is currently possible.
To detect signs of autism in the electrical activity of the brain, a special contrasting variational autoencoder was used. For a more accurate analysis of electroencephalogram data, it was improved using mathematical methods of graph theory.
The new development was able to detect autism with a 95% probability, whereas similar AI-based systems previously did this with 80% accuracy.
The developed autoencoder recorded even non-obvious patterns present in one data set and absent in other image banks or measurement results.
Scientists managed to identify one important property for the diagnosis of autism. It was the weakening of some functional connections in the frontal lobe of the brain.
Earlier www1.ru reported that NtechLab started creating a neural network for the diagnosis of Alzheimer's and dementia.
Read materials on the topic:
AI helped doctors in Moscow make 14 million preliminary diagnoses
A neuropsychiatry laboratory will appear in Tomsk: virtual reality will be used to treat patients