Scientists have developed an algorithm that is many times better at detecting epilepsy on EEG recordings than other automated methods. To do this, the authors combined two approaches to the analysis of brain activity signals — a classifier that does not require training and a learning neural network. The development will make it possible to automate EEG analysis and thereby simplify the process of detecting epilepsy. The results of the study, supported by a grant from the Presidential Program of the Russian Science Foundation, were published in the journal IEEE Access.

Epilepsy is considered one of the most common neurological diseases: it affects about 50 million people worldwide. Epileptic seizures occur due to abnormal activity of various areas of the brain and may be accompanied by loss of consciousness, uncontrolled movements, impaired vision and cognitive abilities. To date, doctors are quite successful in fighting epilepsy — approximately 70 percent of patients with this diagnosis have seizures that stop after drug treatment or surgery.

The only clinical method of diagnosing epilepsy is the detection of its electroencephalographic biomarkers – special patterns on the EEG of patients. It is important to identify them, because not all types of epilepsy are accompanied by seizures, and it is not always possible to make a diagnosis based solely on external symptoms.

However, this is a rather time-consuming process: the data set for one patient can be from ten to several days of recording. In addition, the doctor needs to distinguish the signals characteristic of epilepsy from other types of brain activity, which requires serious training and long-term clinical practice.

Scientists from Immanuel Kant Baltic Federal University (Kaliningrad), M.I. Pirogov Russian National Research Medical University (Moscow) and Immersmed LLC (Moscow) have developed an automated method for detecting brain activity corresponding to epilepsies on EEG recordings. The authors took as a basis two approaches to detecting attacks and combined them, thereby creating a two-stage system.

In the first stage, a simple algorithm called a classifier, which requires training, detected “outliers” in the EEG recordings – signals whose intensity is outside the limits of normal brain activity. Emissions can be both epileptic attacks and various external noises, some episodes of atypical brain activity, for example, sleep spindles during the patient’s sleep. Thus, at the output of the classifier, a markup is obtained, which contains both real epileptic seizures and various false components.

Therefore, in the second stage, the neural network (a more complex algorithm based on machine learning) analyzed in more detail the EEG recordings, which were marked as “suspicious” by the first algorithm, and concluded whether the EEG has epilepsy or not.

The authors used a convolutional neural network, which is often used for image analysis. She considered EEG recordings not as a set of signals, but as a complete image in which she found the necessary signals. In this context, the neural network simulated the work of a doctor who, in his search for an epileptic seizure, also analyzes signals and spectra for the presence of certain patterns.

The researchers tested the proposed two-stage system, as well as both of its elements separately. For this, we used EEG recordings taken from 83 people with epilepsy, during attacks and in a calm state (with normal brain activity).

It turned out that the sensitivity – the ability to detect abnormal signals on the EEG – of the classifier and the neural network separately reaches 90 and 96 percent, respectively. However, the accuracy of these approaches turned out to be quite low – 12 and 13 percent, and this indicates that the algorithms confuse coital epilepsy with other types of abnormal brain activity.

The two-step approach showed a sensitivity of 84 percent, but a much higher accuracy of 57 percent, with fewer false positives. Therefore, it is much better suited for potential application in clinical practice than the approaches included in it separately.

“The obtained result gives hope for the creation of an automatic system of epileptic EEG marking, which will significantly reduce the routine burden of marking many hours of recordings on epileptologists. The proposed marking system is currently being implemented in the form of a software product – an online medical service – by colleagues from Immersmed LLC and can be used in many medical centers in Russia,” says Oleksandr Khramov, Doctor of Physical and Mathematical Sciences, the head of the project supported by a grant from the Russian National Academy of Sciences , professor, chief researcher of the Immanuel Kant Baltic Center of Neurotechnology and Artificial Intelligence of BFU.