Automated interictal epileptic spike detection from simple and noisy annotations in MEG data.
Pauline Mouches1, Julien Jung2,3, Armand Demasson2
1Université Claude Bernard Lyon 1, CNRS, INSERM, Centre de Recherche en Neurosciences de Lyon CRNL U1028 UMR5292, EDUWELL, 69500, Bron, France. pauline.mouches@inserm.fr.
Brain Structure & Function
|May 30, 2026
Summary
Deep learning models can detect interictal epileptic spikes in magnetoencephalography (MEG) recordings, even with imperfect data. This advances automated analysis for drug-resistant epilepsy presurgical evaluations.
Area of Science:
- Neuroscience
- Medical Imaging
- Artificial Intelligence
Background:
- Presurgical evaluation is crucial for drug-resistant epilepsy.
- Magnetoencephalography (MEG) aids in localizing the epileptogenic zone by identifying interictal epileptic spikes.
- Manual spike detection is time-consuming and has moderate inter-rater agreement, while current automated methods lack robustness or require extensive data.
Purpose of the Study:
- To develop and evaluate deep learning models for automated interictal spike detection in MEG recordings.
- To assess model performance using real-world clinical data with limited annotations.
- To explore the utility of interactive machine learning for improving data annotation and model robustness.
Main Methods:
- Two deep learning architectures, a feature-based artificial neural network (ANN) and a convolutional neural network (CNN), were proposed.
- Models were trained and evaluated on a database of 82 patients, using temporal and single-expert annotations.
- An interactive machine learning strategy was employed to iteratively enhance data annotation quality.
Main Results:
- Both the ANN and CNN models outperformed the state-of-the-art model in detecting interictal spikes (F1-scores: CNN=0.46, ANN=0.44).
- The models demonstrated robustness to noisy annotations, indicating suitability for imperfect real-world data.
- The interactive machine learning approach showed potential for accelerating data annotation.
Conclusions:
- Deep learning models are effective for automated interictal spike detection in MEG, even with complex and imperfectly annotated data.
- The proposed models offer efficient tools for clinical presurgical evaluation in drug-resistant epilepsy.
- Interactive machine learning can significantly improve data annotation efficiency and model performance.
