Domain adaptation to automatic classification of neonatal amplitude-integrated EEG

Amplitude-integrated electroencephalographic (aEEG), a method for continuous long-term monitoring of brain activities, is widely used for clinical needs in monitoring newborns. While the variation in aEEG signals from different individuals causes differences in the data distribution, the task to model and automatically classify aEEG signals across different individuals is challenging. In this paper, a domain adaptation algorithm is introduced to the automatic classification of neonatal aEEG signals. The aEEG signal of an individual represents a domain. Signals from multiple training individuals form the source domains and those from test individuals form the target domains. Several auxiliary classifiers are trained and then combined into a robust target classifier. The key feature of the algorithm is a weighting scheme that leverages all the classifiers learnt from the labeled signals across multiple source domains. Experiments on aEEG tracings of 103 cases were conducted to validate the method. The result shows that the domain adaptation method increases the classification accuracy by about 10% compared with the case without domain adaptation method. Besides, the domain adaptation method can reach quite a high accurate rate with only a few training sets. The novel automatic detection of aEEG could be helpful in bedside brain disorder monitoring in newborns.