Improved multi-stage neonatal seizure detection using a heuristic classifier and a data-driven post-processor

•An improved neonatal seizure detection method is discussed.•A set of characteristic features of seizures are identified by data-driven methods.•Described core characteristics of neonatal seizures can easily be used for other automated methods.

ObjectiveAfter identifying the most seizure-relevant characteristics by a previously developed heuristic classifier, a data-driven post-processor using a novel set of features is applied to improve the performance.MethodsThe main characteristics of the outputs of the heuristic algorithm are extracted by five sets of features including synchronization, evolution, retention, segment, and signal features. Then, a support vector machine and a decision making layer remove the falsely detected segments.ResultsFour datasets including 71 neonates (1023 h, 3493 seizures) recorded in two different university hospitals, are used to train and test the algorithm without removing the dubious seizures. The heuristic method resulted in a false alarm rate of 3.81 per hour and good detection rate of 88% on the entire test databases. The post-processor, effectively reduces the false alarm rate by 34% while the good detection rate decreases by 2%.ConclusionThis post-processing technique improves the performance of the heuristic algorithm. The structure of this post-processor is generic, improves our understanding of the core visually determined EEG features of neonatal seizures and is applicable for other neonatal seizure detectors.SignificanceThe post-processor significantly decreases the false alarm rate at the expense of a small reduction of the good detection rate.