Development of a deep learning tool to detect drug-resistance epilepsy with EEG
Epilepsy & Behavior · 2026-07 · 180 · 111028
DOI: 10.1016/j.yebeh.2026.111028 ↗Kurz gefasst
Fachartikel in der Zeitschrift Epilepsy & Behavior, 2026-07. Verfasst von Manuel Alanís-Bernal und 11 weiteren Autor:innen. Beteiligte Einrichtung: Neurology Department, Vall d'Hebron University Hospital, Barcelona, Spain. Hauptthemen laut MEDLINE: Electroencephalography, Drug Resistant Epilepsy, Deep Learning. Die Arbeit nennt 27 Literaturangaben.
Abstract (Originalsprache)
INTRODUCTION: Drug-resistant epilepsy affects about 30% of patients and is linked to poorer outcomes. Deep learning can extract complex patterns from EEG data and shows promise for identification of drug resistance (DRE). This study aimed to develop a convolutional neural network (CNN) to classify patients with epilepsy as drug-resistant or non-drug-resistant using surface EEG. MATERIAL AND METHODS: We conducted a retrospective single-center study including patients with epilepsy recorded between 2020/01-2024/12. Each patient contributed one wake EEG (10-20; ≥30 min), preprocessed and labeled as drug-resistant or non-drug-resistant according to ILAE criteria. Patients with prior epilepsy surgery, progressive etiologies, or developmental and epileptic encephalopathies were excluded. Multiple CNN models were developed by an analyst blinded to all clinical data, using only the EEG images and their corresponding drug-resistance labels for training. Data were split into training-validation (80%) and independent test (20%) sets; tenfold cross-validation with bootstrap resampling optimized performance and avoid overfitting. The final model was evaluated on the test set for real-world applicability. RESULTS: We included 180 patients (mean age 41.1 ± 17 years; 52.8% female), of whom 87 (48.3%) met DRE criteria. Median epilepsy onset was 19 years [IQR 10-35]. DRE patients had longer disease duration and more frequent focal epilepsy. The best model achieved 73.5% validation accuracy and, on the independent test set, 69.4% accuracy (sensitivity 68.4%, specificity 70.6%, F1-score 0.70, AUC-ROC 0.70 and AUPRC 0.66). CONCLUSION: A deep learning model using only surface EEG data showed moderate performance, which could improve by adding neuroimaging and clinical information from larger datasets.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Neurology Department, Vall d'Hebron University Hospital, Barcelona, Spain.
- Velmon Technologies, Barcelona, Spain.
- Epilepsy Unit, Neurology Department, Vall d'Hebron University Hospital, Barcelona, Spain; Research Group on Status Epilepticus and Acute Seizures, Vall d'Hebron Research Institute (VHIR), Vall d'Hebron University Hospit…
Zitieren
APA
Alanís-Bernal, M., Vélez-Rodríguez, J. R., Campos-Fernández, D., Quintana, M., Abraira, L., Fonseca, E., López-Maza, S., Lallana, S., Garcés-Pellejero, M., Bellido-Castillo, E., López-Aylagas, I., & Toledo, M. (2026). Development of a deep learning tool to detect drug-resistance epilepsy with EEG. Epilepsy & Behavior, 180, 111028. https://doi.org/10.1016/j.yebeh.2026.111028
Vancouver
Alanís-Bernal M, Vélez-Rodríguez JR, Campos-Fernández D, Quintana M, Abraira L, Fonseca E, et al. Development of a deep learning tool to detect drug-resistance epilepsy with EEG. Epilepsy & Behavior. 2026;180:111028. doi:10.1016/j.yebeh.2026.111028
BibTeX
@article{AlansBernal2026,
author = {Alanís-Bernal, Manuel and Vélez-Rodríguez, José Ramón and Campos-Fernández, Daniel and Quintana, Manuel and Abraira, Laura and Fonseca, Elena and López-Maza, Samuel and Lallana, Sofia and Garcés-Pellejero, María and Bellido-Castillo, Enric and López-Aylagas, Ignacio and Toledo, Manuel},
title = {Development of a deep learning tool to detect drug-resistance epilepsy with EEG},
journal = {Epilepsy & Behavior},
year = {2026},
volume = {180},
pages = {111028},
doi = {10.1016/j.yebeh.2026.111028}
}
EUVIMED