Two-stage computer algorithm will detect epilepsy with high precision
Scientists elaborated algorithm that much better detects epilepsy on EEG recordings, than other automated methods. To achieve this, authors combined two approaches to analysis of signals of brain activity—classifier, that doesn't require education, and trainable neural network. The project will enable to automate analysis of EEG and so simplify the process of detecting of epilepsy. Results of the research, supported by the grant of Presidential program of Russian Scientific Foundation, are published in the magazine IEEE Access.
Epilepsy is considered to be one of the most widespread neurological diseases: 50 million people all over the world suffer from it. Epileptic seizures occur due to abnormal activity of different areas of brain and can be followed by loss of consciousness, uncontrolled movements, visual and cognitive disorders. Nowadays doctors struggle with epilepsy rather successfully—about 70% patients with such diagnosis after medical therapy or surgical interference experience stop of seizures. In order to make diagnosis precisely and to assign a correct therapy, doctors search signs of epilepsy on EEG recordings. This process is rather laborious: data record for one patient can take from ten to several tens hours of recording. Besides this a doctor needs to distinguish signs that are characteristic for epilepsy, from other kinds of brain activity, that requires strong background and long clinical practice.
Scientists from Immanuel Kant Baltic Federal University (Kaliningrad), Pirogov Russian National Research Medical University (Moscow) and Limited Society "Immersmed" (Moscow) developed an automated method for detection of brain activity, corresponding to epileptic seizures, in EEG recordings. As a base authors took two approaches to detection of seizures and combined them, thus creating a two-stage system.
In frames of the first stage a simple algorithm called classifier which doesn't require training, detected on EEG recordings "emissions"—signals, the intensity of which surpass normal brain activity. Emissions can be both seizures of epilepsy, and also various external noises, some episodes of atypical activity of brain, for example, sleep spindles during patient's sleep. Thus, on the exit of classifier you get a marking, that contains both true epileptic seizures and various false components.
That's why further—on the second stage—neural net (more complicated algorithm on the base of machine learning) studied in more detail those EEG recordings, that were marked by the first algorithm as "suspicious", and made conclusion whether there was really epilepsy on EEG or not.
Authors used neural net of convolutional type, that is often used for analysis of images. It treated EEG recordings not as a set of signals but as entire image, on that it found demanded signals. In that context neural network imitated doctor's work, that in the search of epileptic seizure also analyzes signals and spectra for existence of certain patterns.
Researchers tested the suggested two-stage system and also two its elements separately. For this aim they used EEG recordings from 83 people, suffering from epilepsy, during seizures and calm state (with normal brain activity).
It turned out that sensibility—the ability to detect abnormal signals on EEG—of classifier and neural network separately achieves 90 and 96% correspondingly. However, the proximity of these approaches turned out to be rather low—12 and 13%, and it points to the fact that algorithms confuse epilepsy with other types of abnormal brain activity. Two-stage approach showed sensibility 84%, but much higher—57%—proximity due to the reduction of false positive results. That is why it much more suitable for potential use in clinical practice, than separate approaches that it includes.
"The obtained result promises creation of automated system of marking of epileptic EEG, that enables to reduce routine duties of doctors epileptologists, connected with marking of long recordings, significantly. The proposed system of marking now is realized in the form of program product—online medical service—by colleagues from Limited Society "Immersmed" and can be applied in many medical centers of Russian Federation",—tells the head of the project, supported by the grant of RSF, Alexander Hramov, Doctor of Physical and Mathematical Sciences, professor, leading Researcher at the Center for Neurotechnology and Machine Learning, Immanuel Kant Baltic Federal University.
More information:
doi.org/10.1109/ACCESS.2024.3453039
Provided by Immanuel Kant Baltic Federal University