Comprehensive Summary
This paper describes the usage of interpretable deep learning for automated seizure detection in pediatric electroencephalography (EEG). This was achieved by creating two novel models, the squeeze-and-excitation fully convolutional network (SE-FCN), which was designed to strengthen sensitivity to the most relevant EEG channels, and a transformer-based network (TransNet), which was optimized to capture long-range temporal dependencies and cross-channel relationships. Both proposed models significantly outperformed baseline approaches with the SE-FCN achieving an overall accuracy of 87%, and the TransNet achieving with an accuracy of 86%, demonstrating its stronger ability to generalize temporal information across patients. The models also produced interpretable outputs as heatmaps generated from the squeeze-and-excitation blocks and transformer attention weights consistently highlighted electrodes and brain regions with activity. The researchers categorized patients into five epileptogenic zone groups based on these outputs, demonstrating that the models classify seizures and provide clinical localization.
Outcomes and Implications
This research is important because epilepsy remains one of the most common neurological disorders, with diagnosis and its monitoring reliant on EEG interpretation. However, current CNN and SGN networks lack interpretability and clinical applicability, as they are unable to capture patterns and exhibit significant variability. The researchers aimed to address this limitation by combining high diagnostic accuracy with interpretability, therefore creating tools that can both detect seizures and assist in localizing epileptogenic regions. This is especially significant for children with drug-resistant epilepsy, where accurate localization of seizure onset zones is essential to determining the eligibility for surgical intervention or neurostimulation. The high accuracies of 86–87% suggest that the models are clinically acceptable. However, the authors note that large-scale validation is still required, and computational efficiency needs to be optimized for real-time applications. According to the researchers, once the novel models are successfully adopted, they could reduce diagnostic delays, improve seizure localization, and support more personalized treatment planning, especially in pediatric epilepsy.