Regularization of Deep Neural Networks for EEG Seizure Detection to Mitigate Overfitting

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Seizure detection from EEG records is crucial for clinical workflows.
• Existing algorithms struggle with generalization beyond training data, limiting their effectiveness for broader patient populations.
• Patient-specific EEG patterns often lead to overfitting in deep learning models.

Purpose of the Study:

• To develop a robust data processing pipeline for intra-patient seizure detection using EEG data.
• To create spatially and session-invariant features to improve model generalization.
• To enhance the accuracy and sensitivity of seizure detection algorithms.

Main Methods:

• Implemented a data processing pipeline on an intra-patient EEG seizure dataset.
• Developed spatially and session invariant features by regularizing deep learning networks.
• Utilized random channel rearrangements and data rescaling techniques for regularization.
• Applied adversarial multi-task learning as an additional regularization method.

Main Results:

• Baseline deep learning model without regularization achieved an F1 score of 0.544.
• Random channel rearrangement improved the F1 score to 0.629.
• Random data rescaling further increased the F1 score to 0.651.
• The best model achieved a sensitivity of 0.86, outperforming baseline studies.

Conclusions:

• Patient and session-specific dependencies cause overfitting in deep learning models for EEG seizure detection.
• Regularization techniques, including random rearrangement and rescaling, prevent models from learning patient-specific features.
• The proposed methods significantly improve intra-patient seizure detection accuracy and generalization.