Heuristic Active Learning for the Prediction of Epileptic Seizures Using Single EEG Channel

Predicting epileptic seizure occurrence has long been a goal of the community surrounding it. Accurate prediction, however, is still elusive. This work presents a modified pipeline for the training of seizure prediction systems which aims to attenuate the effects of current data labeling strategies - and consequent data mislabeling of samples that heavily affect classifiers that are trained on it. This paper also presents a seizure prediction system trained following the proposed pipeline, which improved our system's performance by reducing its time-in-warning (TiW) by over 14%, while improving its prediction sensitivity to 72.4%, bringing its performance closer to the state-of-the-art performance (83.1% prediction sensitivity) for systems with similar TiW (41%) [1], while only requiring input from two scalp EEG electrodes - without making use of any variables external to the single EEG channels.