Cracking the Seizure Code: Leveraging Bi-LSTM Models for Neonatal EEG Interpretation and Seizure Classification

This research investigates how Bidirectional Long Short-Term Memory (Bi-LSTM) networks can be used to identify neonatal seizures from EEG data, with an emphasis on mirroring expert neurologist annotations rather than independently predicting seizures themselves.Timely detection of seizures in infants is crucial in Neonatal Intensive Care Units (NICUs); however, traditional methods often delay critical intervention and treatment.Machine learning offers a promising alternative solution.For this study, EEG data from 79 neonates within a publicly available dataset were cleaned, filtered, split into appropriate segments, and processed so the model could learn most effectively.The Bi-LSTM model, trained over eight epochs, achieved an overall accuracy of 83%.A key limitation of this study is the absence of "ground truth," as even expert annotations are subjective, which introduces variation within the training data.Additionally, the overlap between seizure and normal brain activity in the EEG signals contributes to lower model accuracy in certain patient cases.Nevertheless, this study highlights the potential of Bi-LSTM models in enhancing neonatal seizure detection and improving long-term outcomes for vulnerable neonates, offering an essential step toward faster, more effective diagnosis in NICUs.