Seizure Prediction using Convolutional Neural Networks and Sequence Transformer Networks.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate seizure prediction is crucial for developing advanced epilepsy management devices, such as closed-loop neuromodulation systems.
• Nonstationarity in electroencephalogram (EEG) signals presents a significant challenge for reliable seizure prediction models.
• Existing prediction models often struggle with the dynamic and variable nature of epileptic seizures.

Purpose of the Study:

• To present a patient-specific deep learning approach to enhance the accuracy of seizure prediction.
• To address the challenge of EEG signal nonstationarity through data transformation techniques.
• To improve the performance of predictive models for epilepsy management.

Main Methods:

• A Sequence Transformer Network (STN) was employed to learn temporal and magnitude invariances in EEG data.
• Short-time Fourier Transform (STFT) was computed on transformed EEG signals, serving as input features for a Convolutional Neural Network (CNN).
• A k-out-of-n post-processing method was utilized to minimize isolated false positives.

Main Results:

• The proposed model achieved a sensitivity of 82% on the American Epilepsy Society Seizure Prediction Challenge dataset.
• The false prediction rate was reported as 0.38 per hour.
• An average Area Under the Curve (AUC) of 0.746 was obtained, demonstrating robust predictive performance.

Conclusions:

• The patient-specific deep learning approach effectively improves seizure prediction accuracy by transforming EEG data.
• The integration of STN, STFT, and CNN offers a powerful framework for analyzing nonstationary EEG signals.
• This method holds promise for the development of more reliable wearable and implantable devices for epilepsy monitoring and treatment.