Czech Republic, 29 January 2026 – Scientists at IT4Innovations, VSB – Technical University of Ostrava, have developed a new artificial intelligence model capable of detecting epileptic seizures from EEG recordings of brain activity with high accuracy. When tested on clinical data, the system achieved an average accuracy of 96 percent. The research results were published in the Neurocomputing journal.

Epilepsy is one of the most common brain disorders and, according to the World Health Organization, affects more than 50 million people worldwide. In clinical practice, doctors primarily rely on electroencephalography (EEG), which records the brain’s electrical activity. However, analysing these recordings is time-consuming and complex, as the signals are affected by, e.g., noise, and seizures themselves represent only a small fraction of the overall data.

The new epilepsy diagnostic model developed in Ostrava addresses this challenge using a convolutional neural network (CNN) with Dynamic Channel Selection (DCS) and a Feature Generation (FG) module, and is therefore referred to as FG-DCS-CNN. The model can automatically focus on the most relevant signals from EEG electrodes, reduce noise, and identify changes in brain activity that may indicate an ongoing epileptic seizure.

Compared to Transformer-based models, FG-DCS-CNN achieves significantly better results. This is due to its architecture, which is tailored to the spatiotemporal nature of EEG data, and it performs better on smaller datasets than methods that require extensive training.

The model searches EEG data for subtle and complex changes in brain signals – it’s like looking for suspicious fingerprints in brain activity. During training, it decides for itself which signals are most important for detecting a seizure. The model is also computationally efficient, highlighting its potential for use on much larger datasets.

The model was trained and tested on data from patients with epilepsy, which dozens of recorded seizures. For some seizure types, the model’s accuracy reached almost 100 percent. The researchers also observed a very low number of false alarms and missed seizures.

A key advantage of the model is its processing speed. The system can analyse EEG signals within tens of milliseconds, enabling real-time or near-real-time use in clinical practice. It allows doctors not only to evaluate long-term EEG recordings more quickly and accurately, but also to monitor critical moments in patients, plan examinations, and adjust treatment where necessary. The FG-DCS-CNN model acts as a “smart detective” of brain activity: it identifies which signals from the electrodes are crucial for recognising a seizure, filters out noise that could obscure key seizure patterns, and detects even the most subtle changes in brain activity that might be missed by the human eye or conventional methods. This reduces the risk of missed seizures or false alarms and saves doctors valuable time that would otherwise be spent on manual analysis, including patient self-reporting and EEG interpretation.

The model was developed by researchers Muhammad Zeerak Awan, Adil Jhangeer, Petr Strakoš, and Lubomír Říha. Computations and testing were carried out on IT4Innovations’ supercomputers, with support from the LERCO project.

In the future, the team plans to validate the technology using data from additional hospitals and to incorporate modern approaches that enable training models without the need to share sensitive medical information between institutions.