Machine Learning for Short-term Seizure Forecast Using Neonatal EEG.

We propose a novel machine-learning method for continuous short-term seizure forecast by neonatal electroencephalogram (EEG) recorded in the neonatal intensive care unit (NICU). The proposed model was based on an Adaptive Boosting classifier, and supported by 22 features extracted from short-time segmented EEG. The proposed method is trained/tested on three datasets (two open datasets and one privately) recorded from 148 neonates with a diversity of gestation periods (late preterm and full-term) and diagnoses (dominantly present hypoxic-ischemic encephalopathy HIE infants). The achieved performances for Matthews correlation coefficient and area under the receiver-operating characteristic curve, (MCC 0.466±0.078, AUROC 0.738±0.041) respectively, outperformed the current state-of-art ML model (MCC 0.255±0.054, AUC 0.678±0.041).Clinical Relevance- The development of a ML model for predicting short-term seizures can contribute to the optimization of allocated resources for monitoring in the NICU and improve and facilitate decision-making algorithm in treatment of newborns with hypoxic-ischemic encephalopathy. The proposed method achieves state-of-the-art performance for short-term seizure forecast.