DeepIED: An epileptic discharge detector for EEG-fMRI based on deep learning

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate presurgical delineation of the epileptogenic zone (EZ) is crucial for successful epilepsy surgery.
• Electroencephalography-functional Magnetic Resonance Imaging (EEG-fMRI) is a valuable noninvasive tool for EZ estimation.
• Manual marking of interictal epileptic discharges (IEDs) in degraded in-scanner EEG is a significant bottleneck for EEG-fMRI.

Purpose of the Study:

• To develop and validate a semi-automatic deep learning-based detector for IEDs in EEG recorded during fMRI.
• To reduce the manual annotation burden for EEG-fMRI analysis.
• To improve the accuracy and practicality of EEG-fMRI for presurgical epilepsy evaluation.

Main Methods:

• A deep learning model was trained on 30 patients' EEG data to detect candidate IEDs in-scanner.
• The model identifies IEDs resembling externally recorded sample IEDs.
• Validation involved accuracy assessment on 37 patients and reproducibility testing on 15 patients.

Main Results:

• The deep learning method significantly improved median IED detection sensitivity from 50.0% to 84.2% compared to template-based methods.
• The false positive rate was maintained at 5 events/min.
• Concordance between hemodynamic response maps and intracerebral EEG-defined EZ was comparable to manual marking (76.9%).

Conclusions:

• The proposed deep learning-based semi-automatic IED detector substantially enhances EEG-fMRI analysis efficiency and accuracy.
• This tool simplifies IED identification, reducing expert workload and making EEG-fMRI more clinically feasible.
• Improved EZ delineation through this method can lead to better surgical outcomes for refractory epilepsy patients.