Optimizing Seizure Prediction From Reduced Scalp EEG Channels Based on Spectral Features and MAML

Epilepsy is a severe neurological disease with high prevalence and morbidity worldwide. The unpredictability of seizures prevents physicians from tailoring drugs and therapies. Recent non-invasive seizure prediction research has not improved the overall quality of life for patients. Therefore, new research studies on seizure prediction must integrate data, embedded devices, and algorithms. For a seizure prediction system to emerge as a feasible solution, we must address a reduction in EEG scalp electrode channels, along with a decrease in computational resources to train the time-series signal. In this work, we propose an optimized patient-specific channel reduction for seizure prediction using Model Agnostic Meta-Learning (MAML) applied to a Deep Neural Network (DNN). We selected and optimized the number of channels from each of the 23 subjects of the CHB-MIT Dataset. The feature vectors are extracted using Ensemble Empirical Mode Decomposition (EEMD) and Sequential Feature Selection (SFS). We implemented the MAML model to classify the small EEG data generated from the reduced number of subject-dependent electrodes. The experiment results yield an average sensitivity and specificity of 91% and 90%, respectively. Our study demonstrates that MAML is a promising approach to learn EEG patterns to predict epileptic seizures with few EEG scalp electrodes.