Abstract
Over the past few decades, electroencephalography (EEG) monitoring has becomea pivotal tool for diagnosing neurological disorders, particularly fordetecting seizures. Epilepsy, one of the most prevalent neurological diseasesworldwide, affects approximately the 1 \% of the population. These patientsface significant risks, underscoring the need for reliable, continuous seizuremonitoring in daily life. Most of the techniques discussed in the literaturerely on supervised Machine Learning (ML) methods. However, the challenge ofaccurately labeling variations in epileptic EEG waveforms complicates the useof these approaches. Additionally, the rarity of ictal events introduces anhigh imbalancing within the data, which could lead to poor predictionperformance in supervised learning approaches. Instead, a semi-supervisedapproach allows to train the model only on data not containing seizures, thusavoiding the issues related to the data imbalancing. This work proposes asemi-supervised approach for detecting epileptic seizures from EEG data,utilizing a novel Deep Learning-based method called SincVAE. This proposalincorporates the learning of an ad-hoc array of bandpass filter as a firstlayer of a Variational Autoencoder (VAE), potentially eliminating thepreprocessing stage where informative band frequencies are identified andisolated. Results indicate that SincVAE improves seizure detection in EEG dataand is capable of identifying early seizures during the preictal stage as wellas monitoring patients throughout the postictal stage.