Engineering nonlinear epileptic biomarkers using deep learning and Benford's law.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Early seizure detection in intracranial EEG is crucial for patient management.
• Traditional engineered metrics have limitations in capturing complex seizure dynamics.

Purpose of the Study:

• To design and evaluate deep neural networks for early seizure detection.
• To compare deep learning features with established engineered metrics.
• To identify significant biomarkers for seizure prediction.

Main Methods:

• Two deep neural networks were designed to encode 16 features from intracranial EEG.
• Models were pretrained for seizure identification using leave-one-out cross-validation on 127 seizures.
• Feature spaces were interpreted using spectral power modulations and Random Forest Classifiers (RFCs).

Main Results:

• Deep learning models extracted unique feature spaces, outperforming or complementing engineered metrics.
• Most significant features (MSFs) were identified using Gini Importance from RFCs.
• A transferable method achieved an AUC score of 0.93 in a Kaggle challenge.

Conclusions:

• Deep neural networks offer a powerful approach for early seizure detection.
• The identified biomarkers are transferable and enhance seizure prediction accuracy.
• This method provides interpretable insights into seizure dynamics.