Epilepsy is one of the most common neurological disorders, affecting 50 million people globally. Epileptic seizures result from abnormal brain activity and can lead to loss of consciousness, uncontrolled movements, and various visual and cognitive impairments. Currently, about 70% of epilepsy patients experience cessation of seizures through medical therapy or surgical intervention. To diagnose epilepsy accurately and prescribe the right treatment, doctors rely on identifying epileptic signs in EEG recordings. This process, however, is time-consuming, as EEG recordings for a single patient can range from hours to days. Furthermore, distinguishing epilepsy-specific signals from other brain activity requires significant expertise and clinical experience. Now, scientists have developed an algorithm that surpasses existing automated methods in detecting epilepsy on EEG recordings. The approach combines two techniques—an unsupervised classifier and a trainable neural network—aiming to automate EEG analysis and simplify the epilepsy detection process.

A team of scientists, including researchers from Immanuel Kant Baltic Federal University (Kaliningrad, Russia), has created an automated method to detect brain activity corresponding to epileptic seizures in EEG recordings. They employed a two-stage detection system by integrating two different approaches. In the first stage, a simple, unsupervised algorithm, known as a classifier, identifies "emissions"—signals with intensities that exceed normal brain activity. These emissions can represent epileptic seizures, external noise, or atypical brain activity such as sleep spindles. The classifier produces a marking that includes both actual epileptic seizures and various false positives.

In the second stage, a more complex neural network, based on machine learning, examines the "suspicious" EEG recordings flagged by the classifier. This neural network, specifically a convolutional type often used in image analysis, evaluates the EEG data as an entire image rather than as individual signals, identifying patterns associated with epilepsy. In this way, the neural network mimics how a doctor examines EEG signals and spectra for specific markers of epileptic seizures. The researchers tested both the two-stage system and its individual components using EEG data from 83 epilepsy patients during seizure episodes and calm states.

According to results published in IEEE Access, the sensitivity—the ability to detect abnormal EEG signals—of the classifier and neural network individually was 90% and 96%, respectively. However, their specificity, or ability to differentiate epileptic activity from other types of abnormal brain signals, was low at 12% and 13%. The two-stage system, while slightly less sensitive at 84%, had a much higher specificity of 57%, indicating a significant reduction in false positives. This makes the combined approach more suitable for clinical use than either method alone.

“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,” said Alexander Hramov, head of the project and researcher at Immanuel Kant Baltic Federal University. 

http://www.gzjiayumed.com/en/index.asp .