EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

EEG-Based Cross-Patient Epileptic Seizure Detection via State-Conditional Domain Adversarial Training and Temporal Modeling

Rina Tazaki, Tomoyuki Akiyama, Akira Furui

IEEE Access · 2026 · 14 · 78191-78207

DOI: 10.1109/access.2026.3694985 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Rina Tazaki und 2 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Japan Society for the Promotion of Science (JSPS) KAKENHI.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Automated epileptic seizure detection from scalp electroencephalogram (EEG) remains challenging due to significant inter-patient variability. While domain adversarial training has emerged as a promising approach for patient-invariant features, existing methods typically rely on convolutional neural network (CNN)-based models that process each EEG window independently, failing to capture temporal progression patterns of seizures. Moreover, conventional domain adversarial training enforces patient invariance uniformly without distinguishing between seizure and non-seizure states, despite inter-patient variability differing across these states. To address these limitations, we propose a domain generalization framework that integrates state-conditional domain adversarial training with temporal modeling via a CNN and bidirectional long short-term memory (BiLSTM), using only source patient data without any target patient enrollment. The two objectives are decoupled into separate stages to avoid conflicting optimization. In the first stage, the CNN extracts state-adaptive patient-invariant features by conditional domain adversarial training. In the second stage, the BiLSTM captures temporal dependencies of these invariant features to model seizure evolution.We evaluated our method using leave-one-patient-out cross-validation on two focal epilepsy datasets: 14 patients from the Siena Scalp EEG Database (long-term recordings) and 20 from a clinical dataset (short-term recordings). The results demonstrate that both conditional domain adversarial training and temporal modeling contribute to performance improvement, with the benefit of state-adaptive invariant feature extraction particularly pronounced in long-term recordings containing diverse background EEG activities. This approach enables immediate application to unseen patients without any enrollment period, providing a clinically practical solution for automated seizure monitoring.

Förderung

  • Japan Society for the Promotion of Science (JSPS) KAKENHI

Zitieren

APA

Tazaki, R., Akiyama, T., & Furui, A. (2026). EEG-Based Cross-Patient Epileptic Seizure Detection via State-Conditional Domain Adversarial Training and Temporal Modeling. IEEE Access, 14, 78191-78207. https://doi.org/10.1109/access.2026.3694985

Vancouver

Tazaki R, Akiyama T, Furui A. EEG-Based Cross-Patient Epileptic Seizure Detection via State-Conditional Domain Adversarial Training and Temporal Modeling. IEEE Access. 2026;14:78191-78207. doi:10.1109/access.2026.3694985

BibTeX
@article{Tazaki2026,
  author = {Tazaki, Rina and Akiyama, Tomoyuki and Furui, Akira},
  title = {EEG-Based Cross-Patient Epileptic Seizure Detection via State-Conditional Domain Adversarial Training and Temporal Modeling},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {78191--78207},
  doi = {10.1109/access.2026.3694985}
}
Kostenlos starten →