Automated seizure detection for neonatal EEG

Epileptic seizures damage the developing brain. Therefore, newborns with high risk of seizures must be continuously monitored to ensure that seizures are detected as quickly as possible. Reliable detection of seizures requires multichannel electroencephalography recordings to be analyzed by experienced clinical experts, such as neurophysiologists. Since this expertise is not always available, there is a demand for automatic seizure detection methods. The challenge in the development of neonatal seizure detection algorithms has been threefold: 1) to define features that best represent the characteristics of neonatal seizures, 2) to find a comprehensive suite of measures to assess the goodness of the algorithm and define its potential utility in clinical practice, and 3) to overcome the lack of open-access data sets that hinders algorithm development. In this thesis, features were defined for detecting the non-stationary periodic characteristics of neonatal seizures in both time and joint time–frequency domains. These features, after excluding outliers, were integrated by a support vector machine incorporated in a neonatal seizure detection algorithm. This algorithm was compared to alternative methods of seizure detection, such as the deep convolutional neural network. To accurately assess algorithm performance, measures for generalizability, non-inferiority to the annotation of the human expert, and clinical relevance were developed. Generalizability was assessed as the consistency of performance between training (through cross-validation) and unseen independent data. Non-inferiority was estimated by comparing the inter-observer agreement of human experts only to a composite human expert/algorithm agreement. Clinical relevance measures evaluated clinically relevant interpretations of algorithm annotation such as short-term seizure burden. The developed measures proved highly discriminatory for seizures, the final algorithm generalized to unseen data, was non-inferior to human annotation with increased diversity in the training data set and provided a clinically relevant interpretation of the recordings. These performance measures comprehensively assess the goodness and clinical efficacy of the algorithm. The data set used to train the proposed algorithm as well as the original algorithm have been made openly accessible to facilitate seizure detector development and comparisons in the future.