Deep learning based automatic detection and dipole estimation of epileptic discharges in MEG: a multi-center study

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Magnetoencephalography (MEG) is vital for diagnosing focal epilepsy.
• Manual identification of epileptic spikes in MEG data is time-consuming for neurophysiologists.

Purpose of the Study:

• To enhance an existing deep learning model for automated epileptic spike detection.
• To validate the multi-center performance of the automated detection model.

Main Methods:

• Developed and trained a deep learning model using data from four MEG centers.
• Evaluated the model on data from two additional independent MEG centers.
• Employed a five-fold subject-wise cross-validation strategy.

Main Results:

• The multi-center model achieved high performance (ROC-AUC 0.9929 internal, 0.9426 external).
• Median dipole detection distances were 4.36 mm (internal) and 7.23 mm (external).
• The multi-center model demonstrated superior performance compared to a single-center model.

Conclusions:

• Multi-center training significantly improves automated epileptic spike detection accuracy.
• The developed model reduces neurophysiologist workload by automating spike analysis.
• This approach shows potential for detecting spikes within 1 cm of manual analysis, saving significant time.