Improvement and Validation of an Automated Neonatal Seizure Detector.

We present the improvements made to and subsequent validation of an automated approach to detect neonatal seizures. The evaluation of the algorithm has been performed on a new and extensive data set of neonatal EEGs. Previously, we have classified neonatal seizures visually into two types: the spike train and oscillatory type of seizures and developed two separate algorithms that run in parallel for their automated detection. The first algorithm analyzes the correlation between high-energetic segments of the EEG, whereas the second one detects increases in low-frequency activity (<8 Hz) and then uses an autocorrelation. An improved version of our automated system (called 'NeoGuard') uses more informative features for classification and optimized parameters for thresholding. The validation was performed on 756 hours of 'unseen' continuous EEG monitoring data from 24 neonates with encephalopathy and recorded seizures. The seizure detection system showed a median sensitivity of 86.9 % per patient, positive predictive value (PPV) of 89.5 % and false positive rate of 0.28 per hour. The modified algorithm has a high sensitivity combined with a good PPV whereas false positive rate is much lower compared to the previous version of the algorithm.