Exploration of Short-range Neonatal Seizure Forecasting with Quantitative EEG Based Deep Learning

Abstract

ABSTRACTBackgroundIn this study, we utilize robust feature selection of quantitative encephalography (QEEG) features for inclusion into a deep learning (DL) model for short-range forecasting of neonatal seizure risk.MethodsWe used publicly available EEG seizure datasets with a total of 132 neonates. The Boruta algorithm with Shapley values was used for QEEG feature selection into a convolutional long short-term memory (ConvLSTM) DL model to classify preictal versus interictal states. ConvLSTM was trained and evaluated with 10-fold cross-validation. Performance was evaluated with varying seizure prediction horizons (SPH) and seizure occurrence periods (SOP).ResultsBoruta with Shapley values identified statistical moments, spectral power distributions, and RQA features as robust predictors of preictal states. ConvLSTM performed best with SPH 3 min and SOP 7 min, demonstrating 80% sensitivity with 36% of time spent in false alarm, AUROC of 0.80, and AUPRC of 0.23. The model demonstrated ECE of 0.106, consistent with moderate calibration. Evaluation of forecasting skill with BSS under varying SPH demonstrated a peak BSS of 0.056 and a trend for decreasing BSS with increasing SPH.ConclusionStatistical moments, spectral power, and recurrence quantitative analysis are predictive of the preictal state. Short-range neonatal seizure forecasting is feasible with DL models utilizing these features.