Romney Díaz, Aníbal

Profile Picture

Filters

20231
Reset filters

Settings

Citations

Publication Search Results

Now showing 1 - 1 of 1
  • Thumbnail Image
    Publication
    A comparison framework for optimizing seizure prediction from reduced scalp EEG channels based on spectral features and DNN meta-learning
    (2023-05-12) Romney Díaz, Aníbal; Manian, Vidya; College of Engineering; Vega, José F.; Juan, Eduardo J.; Torres García, Wandaliz; Other; Pérez Muñoz, Ferando
    Epileptogenesis and unprovoked recurrent seizure are the main obstacles posed in the study of epilepsy. Recent studies are focusing on non-invasive methods for the prediction of seizure onset. Limitations of most seizure prediction methods include a need for a reduced and personalized scalp electrode selection, a model algorithm capable of training with small data samples, and more flexible computational resources. To tackle these limitations, this study presents a patient-specific reduced scalp channel selection for seizure prediction based on model-agnostic meta-learning (MAML). MAML is used to optimize a deep neural network (DNN) trained on a reduced and personalized set of electrodes. The implemented MAML prediction model learns patterns in the preictal and ictal states from each selected group of subject-dependent electrodes. Sequential feature selection (SFS) and empirical ensemble mode decomposition (EEMD) are used to select and extract feature vectors from the most significant electrodes. The EEMD extracted all intrinsic mode oscillatory functions (IMFs) in the segmented frame in the preictal and interictal states. Sixteen IMFs per channel were obtained, and Relief filtered the IMFs with the most significant features to fit the prediction model at different horizon times. The power spectral density in the preictal region exhibits higher values than in the interictal transition, confirming the presence of signature patterns of the preictal stage. These distinctive values in the IMF oscillations were consistent across all horizon times, resulting in a small number of gradient updates for an optimized model. The experiment results yield an average sensitivity and specificity of 91% and 90%, respectively. The False Positive Rate per hour (FPR/h) over three-horizon times was measured as 0.26. This work demonstrates that the proposed non-invasive method represents a compelling alternative to reducing the number of channels in the scalp EEG with a horizon time from 5 minutes to 1 hour.