The neonatal electroencephalogram (EEG) is often contaminated by artifacts. These artifacts make visual inspection difficult and negatively influence the results of automated analysis. There is a significant lack of comprehensive artifact detection systems for the neonatal EEG. We present an automated artifact detection system based on a semi-supervised Gaussian mixture model (GMM). We examined the effects of feature set size, mixture number and the use of principal component analysis (PCA) as a pre-processor. Performance was assessed using the area under the receiver operating characteristic (AUC) and estimated using leave-one-patient-out cross-validation. The best performing system was obtained with 23 features, 30 mixtures and no PCA (median AUC=0.950, IQR: 0.831-0.993). EMG and movement artifacts were detected with the highest accuracy.

Introduction

The neonatal electroencephalogram (EEG) is an important tool for assessing cortical activity in critically ill neonates \cite{hrachovy1990electroencephalography}. By visual interpretation, one can detect abnormalities in the EEG that correlate with abnormal neurological events and long term neurodevelopmental outcome. The interpretation of the EEG is time-consuming and requires significant training. Furthermore, it is not available on demand in many neonatal intensive care units (NICUs), and automated methods of interpretation have been developed to address this problem.
Automated algorithms such as sleep stage detectors \cite{koolen2011sleepclassification} and seizure detectors \cite{stevenson2012nonparametric} are developed to analyze neonatal EEG. However, the EEG is often corrupted by artifacts from internal and external sources making visual inspection more difficult. Internal artifacts, such as ocular and cardiac artifacts, are generated from the variety of body activities of the patient and external artifacts arise from the surrounding environment of the patient. These artifacts are, in principle, present in every EEG recording and reduce the quality of the EEG \cite{griesmaier2011need}. The contamination of the EEG is a major problem because artifacts can resemble cortical activity which can affect the outcome of the automated signal analysis and lead to misinterpretation and misdiagnosis \cite{hagmann2006artifacts,suk2009amplitude,marics2013prevalence}.
The literature presents many artifact removal methods \cite{shao2009automatic,croft2000removal,jung2000removal,jung2000removing,nolan2010faster,gasser2005correction,barlow1986automatic} and artifact detection algorithms \cite{durka2003simple,delorme2007enhanced,gorecka2015detection,delorme2004eeglab,inuso2007brain,vigario1997extraction,schlogl2007fully} in adults, many of which are only applicable to certain artifact types. Especially, several methods of removing muscular artifacts \cite{crespo2008muscle,gasser2005correction,jung2000removing} or ocular artifacts \cite{jung2000removal,croft2000removal,joyce2004automatic} have been proposed. These solutions are not comprehensive since they only focus on specific artifacts. However, there are also artifact detection algorithms with wider scope \cite{durka2003simple,delorme2007enhanced,delorme2004eeglab}. The simplest artifact detection method is based on thresholding \cite{barlow1986automatic}. The major disadvantage of the thresholding methods is that they do not detect artifacts with small amplitudes. Furthermore, the EEG has an enormous array of amplitudes that vary with the age of the patient, recording setup and pathology. Meng and colleagues used a Gaussian mixture model (GMM) in an artifact detection method based on polysomnographic recording of adults in order to improve seizure detection rate \cite{meng2004gaussian}. Durka and colleagues introduced a classifier-based method to remove ocular and power-line interference in polysomnographic recording \cite{durka2003simple}. They only used single features to detect different artifact types, which is often inadequate to distinguish between artifact and non-artifact case. Independent Component Analysis (ICA) has also successfully been implemented in methods of removing artifacts \cite{jung2000removal,vigario1997extraction}. However, removal algorithms are not preferable for medical diagnosis for two reasons 1) diagnosis should not be performed on altered data or data where the confidence of cortical activity being manifest is reduced and 2) long duration recordings are common place so artefactual periods can readily be ignored. In order to effectively execute an automated artifact detection system, the algorithm should be able to deal with all kinds of characteristics found in the EEG signal.
Even though there are several artifact detection and removal algorithms, little research has been done relating to artifact detection in the neonatal EEG. Stevenson and colleagues proposed a neonatal artifact detection method called General Artifact Detection System based on a discriminative classifier (Support Vector Machine) \cite{stevenson2014artefact}. This supervised learning scheme uses labelled data to define class boundaries within the feature space. The labelling of artifacts is, however, not without difficulties: 1) it is a time-consuming process which limits the amount of data available for several artifact classes and 2) the subjectivity of labelling is potentially high due to the number of different artifact types and their perceived influence on EEG interpretation.
In this thesis, the difficulties related to artifacts on EEG are addressed using a GMM-based classifier. This unsupervised learning method defines different classes of the EEG which are then assigned either as EEG from cortical activity or a specific artifact type based on the labelling of the human expert. I investigated the use of different features, the use of principal component analysis for dimensionality reduction and the number of mixtures on artifact detection performance. The performance of the developed system was measured using the area under the receiver operating characteristic estimated using leave-one-patient-out cross-validation. A general introduction to the EEG is presented in Section 2. In Section 3, the research material and methods used in this study are introduced. Results are presented in Section \ref{section4} and discussion in Section \ref{section5}.

Results

This section is only included in papers that rely on primary research. This section catalogues the results of the experiment. The results should be presented in a clear and unbiased way. Most results sections will contain links and citations, e.g., \cite{Feynman_1986}, and equations, e.g. \(e^{i\pi}+1=0\).

Conclusion

The conclusion should reinforce the major claims or interpretation in a way that is not mere summary. The writer should try to indicate the significance of the major claim/interpretation beyond the scope of the paper but within the parameters of the field. The writer might also present complications the study illustrates or suggest further research the study indicates is necessary.