Automated artifact removal as preprocessing refines neonatal seizure detection

Results A significant decrease in false alarms ( p = 0.01) was found while the Good Detection Rate (GDR) for seizures was not altered ( p = 0.50). Conclusions The techniques reduced the number of false positive detections without lowering sensitivity and are beneficial in long term EEG seizure monitoring in the presence of disturbing biological artifacts. Significance The proposed algorithms improve neonatal seizure monitoring. Keywords Neonate Electroencephalography (EEG) Artifact removal Independent Component Analysis (ICA) Seizure Seizure detection 1 Introduction Neonatal seizures indicate serious underlying neurological dysfunction ( Lanska and Lanska, 1996; Volpe, 2008 ). They are caused by spontaneous, excessive rhythmic electrical activity of groups of neurons in response to disturbances in brain function as in metabolic encephalopathies, hypoxia–ischemia, stroke or infections. Standard practice is to treat these seizures with antiepileptic drugs, hoping to prevent additional brain injury ( Cherian et al., 2009 ). Many neonatal seizures are subclinical and are identified only by the labor-intensive technique of continuous EEG monitoring or cEEG ( Hellstrom-Westas et al., 1985; Murray et al., 2008 ). Reliable detection of the seizures during cEEG is essential to enable appropriate treatment. Recently, we have developed an automated EEG-based neonatal seizure detection method ( Deburchgraeve et al., 2008 ). Unfortunately, some artifacts have similar morphology as seizures, and are characterized by a high degree of repetitiveness. Because of this, they are the main cause for false positives ( Fig. 1 ), a common problem in detection algorithms (see also Temko et al. (2011a,b) ). It should be clear that a reliable and clinically usable algorithm should be able to differentiate seizures from artifacts. In this work, we focus on the automated removal of the most common artifacts to reduce the number of false positive detections of our neonatal seizure detector algorithm. We also take the opportunity to discuss some updates of our seizure detector algorithm in Appendix A. Artifacts can originate from various biological and non-biological sources. Commonly present examples are ECG spike artifacts, blood vessel pulsation artifacts, respiration artifacts, electrode artifacts, tremor artifacts, eye artifacts, movement artifacts, artifacts of nearby equipment, etc. As the first three types are – in our experience – the main cause of false positives and relatively independent of the specific monitoring environment, we focus in this work on these artifacts. It has to be emphasized that some artifacts are difficult to differentiate from seizures without extra information. During visual assessment of the neonatal EEG, additional polygraphy signals such as ECG, respiration, movement sensors, etc., are often recorded and required to support the clinical neurophysiologist in recognizing the artifacts ( Cherian et al., 2009 ). The human observer recognizes the artifacts by identifying the time-relation of abnormalities in the EEG channels with the additional polygraphy signals. As shown below, we integrate such criteria of similarity as a preprocessing step for automated seizure detection. EEG signals on the sensor level are a mixture of neuronal activities from multiple cerebral regions and artifactual signals from extracerebral sources because of volume conduction through cerebrospinal fluid, skull, and scalp. All these contributing sources have overlapping scalp projections, time courses, and spectra. Therefore, their distinctive features cannot be separated by simple averaging or spectral filtering, and more advanced techniques need to be used. A common way of separating artifacts from brain signals in adult EEG is based on Blind Source Separation (BSS) techniques. BSS refers to a family of statistical methods that decompose recordings of mixed signals into a number of underlying sources without specifying a priori a specific model. BSS estimates an unmixing matrix that linearly decomposes or unmixes the multichannel scalp data into a sum of temporally independent and spatially fixed components. Different BSS techniques have been used in the literature to extract components of interest ( De Vos et al., 2007 ) or to identify artifacts ( Choi et al., 2005; Joyce et al., 2004; Nam et al., 2002; De Vos et al., 2010, 2011 ). Another well-known BSS method is Second Order Blind Identification (SOBI; Belouchrani et al., 1997 ). SOBI solves the BSS problem using decorrelation across several time points as its basic computational step. This method has proven to be powerful in separating EEG from Electro-oculogram (EOG) sources. In the following, we introduce a general strategy for artifact removal in the neonatal EEG using BSS. Thereafter, we refine the different aspects with respect to the specific needs for the removal of each type of artifact. Finally, we illustrate the potential of the approach by comparing the results of automated seizure detection with and without the artifact removal, using visually assessed and scored EEGs as gold standard. 2 Materials and methods 2.1 Data set All EEG data were recorded at the Sophia Children’s Hospital (part of the Erasmus University Medical Center Rotterdam, the Netherlands) and consisted of continuous recordings of 8 h duration in each of 13 patients. All patients were term neonates with perinatal asphyxia with recorded electrographic seizures. Each measurement started within 24 h of birth. Six of the 13 recordings were made with a full set of 17 EEG electrodes (Fp1,2, F3,4, C3,4, Cz, P3,4, F7,8, T3,4, T5,6, O1,2) placed according to the 10–20 International System, while in the other a reduced set (P3,4 and F3,4 missing) was used. We previously established that the inter-rater agreement between two experienced clinical neurophysiologists in scoring the number of seizures in 1-h segments of EEGs from 10 patients was 73% ( Kappa 0.4). The 8-h segments of EEGs were scored for seizure activity by one of these raters. Seizures were scored when they showed clear variation from the background EEG activity lasting for at least 10 s and showed evolution in time and/or change in amplitude and frequency. Nine datasets were chosen because the original seizure detector had produced several false positives in them (to investigate the potential of applying artifact reduction methods) and four datasets were chosen based on the absence of false alarms (to investigate the effect if any on the good detection rate). Before analysis, the data were filtered between 0.3 and 30 Hz. 2.2 A general BSS strategy for artifact removal The EEG is first decomposed using an appropriate BSS method into its prominent underlying sources ( Fig. 2 A). As different BSS algorithms are available, we need to select an appropriate one for each type of artifact (e.g. De Clercq et al., 2006; Cardoso and Souloumiac, 1993 ). It is difficult to evaluate the quality of various BSS algorithms because we have no direct measurement of the source signals itself. As selection criterion, we therefore decided to use the BSS algorithm that estimates the sources based on assumptions that are valid for the specific type of artifact. One or more of these sources may represent an artifact source and needs to be identified. For this purpose, each source is compared with the simultaneously recorded polygraphy signal that is relevant for the artifact (e.g., the respiration signal for the detection of the respiration artifact). Correct identification of the artifact component is crucial. If the detection of components containing artifact is too sensitive, too many sources may be omitted, resulting in data loss and even missed seizures. If it is not sensitive enough, artifacts may remain in the EEG, resulting in false positive seizure detections. An important problem for this identification is that the shape of the BSS source artifact can be remarkably different from the shape of the signal in the polygraphy measurement ( Strobach et al., 1994 ). Therefore, a direct comparison with the polygraphy signals is not possible. For instance, in case of an ECG spike artifact, the QRS complex of the ECG measurement might only be a small bump in the amplitude of the EEG and thus have a very different morphology. To overcome this problem, we add an artifact-specific processing step before the comparison of the sources with the polygraphy signal. This step transforms the sources and the polygraphy signal in order to enhance the similarity between them ( Fig. 2 A). Once the artifact components are correctly identified, the EEG can be reconstructed without the artifactual source, providing an artifact free EEG. 2.3 ECG spike artifact removal The first step for the removal of the artifacts is the estimation of the sources. For the ECG spike artifact, we use RobustICA. RobustICA is based on the normalized kurtosis contrast function, which is optimized by a computationally efficient gradient-descent technique ( Zarzoso and Comon, 2010 ). It extracts the components one by one using a deflation approach. The assumption of RobustICA is that the components are statistically independent, an approach that works well to separate non-Gaussian sources. Considering that the ECG artifact is strongly super-Gaussian due to its spiky nature, a good choice of ICA algorithm is RobustICA. For the detection of artifactual components, the calculation of the direct correlation between the ICA-components and the simultaneously recorded ECG is not optimal because the morphologies of both signals are very different. To address this problem, we calculated the correlation between the thresholded energy of the ECG with the thresholded energy of the ICA components. For this purpose, we used a non-linear energy operator ψ (NLEO) ( Kaiser, 1990 ). The key property of the operator is: (1) ψ [ A · cos ( ω 0 n + φ ) ] = 1 / 2 · ω 2 · A 2 This indicates that the output ψ of the operator is proportional to the square of both the amplitude A and the frequency ω of the signal. By applying this operator to the ECG and the estimated components, the signal during the presence of a QRS complex will be amplified relative to the background EEG ( Fig. 2 B). Subsequently, the correlation between each energy transformed ICA component and the ECG was calculated. If this correlation was higher than 0.6, a component was identified as artifactual and the EEG was reconstructed without this component. 2.4 Blood vessel pulsation removal RobustICA is not an adequate BSS algorithm for removal of the blood vessel pulsation artifact. A pulsation artifact is a continuous oscillatory type of artifact, highly correlated in time but uncorrelated to the other activity in the EEG, and more Gaussian than the spike train of the ECG spike artifact. The assumption of statistical independence based on higher order moments as used by RobustICA is thus suboptimal. Therefore, we have used Second Order Blind Identification (SOBI) as the BSS algorithm for the problem of pulsation artifact removal. SOBI ( Belouchrani et al., 1997 ) uses second order statistics to decompose the measurements and has been employed in previous EEG studies ( Tang et al., 2004; Vanderperren et al., 2010 ). It is based on a joint diagonalization of correlation matrices. SOBI is most appropriate for sources that are individually correlated in time, but mutually uncorrelated. SOBI considers the relationship between component values at different time lags and decorrelates these values as much as possible. Thus, SOBI uses correlations across time in performing the signal separation. This means that SOBI can isolate highly temporally correlated sources, something that most ICA algorithms cannot do ( Joyce et al., 2004 ). Therefore, it is particularly well-suited for extracting a sine-shaped pulsation artifact. Because the pulsation is synchronized to the heart’s pump cycle, we can use the simultaneously recorded ECG signal to identify source signals containing the pulsation artifact. We then filtered the ECG signal with a low-pass filter at two times the frequency of the heart rate to get rid of the spiky component. The ECG signal now better resembles the pulsation artifact ( Fig. 2 C), enabling a direct correlation and identification of the artifact source. To take the time delay between the ECG and the pulsation artifact into account, we determined the maximum value of the cross-correlation between the filtered ECG and the BSS-sources. If this correlation was higher than 0.4, the EEG can be reconstructed without this source. 2.5 Respiration artifact removal In the previous section about blood vessel pulsation removal, we used SOBI as a tool to estimate the artifact source. The respiration artifact shares many morphological features with the blood vessel pulsation artifact as it is also an oscillatory signal with high autocorrelation. Accordingly, SOBI appears to be also adequate to extract this signal from the EEG. For the respiration artifact removal, the estimated sources were compared to the signal generated by a movement sensor on the abdomen of the baby measuring the respiration. To enhance the similarity between the sources and the reference, both were low-pass filtered at 9 Hz. If the correlation between the source and the reference was higher than 0.4, the source was removed during reconstruction of the EEG. 2.6 Number of BSS components A last parameter that has to be defined for non-deflationary BSS approaches like SOBI is the number of sources that has to be estimated. In general, it is not possible to say how many sources are present in the EEG. Yet, this number is known to affect the quality of the ICA solution. When more sources are estimated than there really are, ICA algorithms tend to ‘overfit’, which can lead to distortions ( Hyvärinen et al., 2001 ). A generally accepted method to define the number of active sources is based on a PCA decomposition of the EEG ( James and Hesse, 2004 ). Then the number of dominant eigenvalues provides an estimate of the number of active sources in the EEG. More specifically, we defined that the number of eigenvalues that explain 95% of the variance is equal to the number of active sources and hence, the number of sources SOBI must estimate. In this way, the number of sources to be extracted is automatically adapted to the complexity of the EEG and the number of EEG channels available. 2.7 Artifact removal as a preprocessing step in automated seizure detection The automated seizure detector works on non-overlapping windows with a length of 30 s. The automated artifact removal presented in this paper is applied to these 30 s windows as a preprocessing step. The number of time points of n -channel data used in the decomposition must be sufficient to obtain the n 2 weights of the unmixing matrix of the BSS. In the literature, it is recommended to have at least some multiple k of these n 2 data points ( Onton et al., 2006 ). For a 17-electrode EEG in a bipolar montage of 20 EEG channels and a complex EEG pattern, the absolute minimum number of samples ( k = 1) is: 1 * 20 2 = 400 samples. At a sampling frequency of 256 Hz, this leads to a minimum window size of 1.6 s. For complex EEG patterns, with many active sources, k can be as high as 20 leading to a minimum window duration of 30 s. In our application k is expected to always remain lower than 20. Hence, the 30 s windows supply the BSS algorithms with sufficient samples to reliably separate the artifact signal from the other concurrent EEG activity and are even adequate for a relatively slow signal such as the respiration. Because we intend to use the automated seizure detector at the bedside, the artifact removal algorithm requires a real-time implementation. The computational aspects of the BSS algorithms are important and may be a limiting factor in the selection of the best BSS algorithm. At present, to ensure real-time operation, RobustICA was limited to 1000 iterations and for SOBI; the number of correlation matrices to be diagonalized was set to 256 samples (or 1 s with a 256 Hz sampling rate). 2.8 Performance measures Assessing the performance of a seizure detection algorithm is not straightforward. Different ways to measure the performance of an algorithm have been reported in the literature, leading to varying results that cannot easily be compared. Recently, Temko et al. (2011b) proposed a naming convention for comparing the performance of seizure detectors, which is followed in this work. In the context of the present study, the most important measure of these is the number of False Alarms per hour (FA/h). This measure directly represents the practical usability of the algorithm, because each FA implies that somebody in the NICU will have to check the patient and the raw EEG recording unnecessarily. Additionally, we also mention the Good Detection Rate (GDR), defined as the number of seizures detected divided by the total number of seizures present. This will illustrate whether seizure activity is incorrectly removed by the artifact reduction methods. In order to assess the statistical significance of the proposed method, we computed a Wilcoxon signed rank test on the number of FA/h and on the GDR with and without Artifact Removal. 3 Results Fig. 3 shows an example of ECG spike train artifact removal. From Fig. 3 B it can be seen that the artifact removal does not alter the other sharp activity in the EEG. Fig. 4 illustrates the joint application of the pulsation and respiration artifact removal on seizure EEG. Note that the artifact removal does not affect the seizure activity ( Fig. 4 A vs. C), despite the similar morphology of the seizure and the artifacts. The results of the seizure detection without and with artifact removal are presented respectively in Tables 1 and 2 . Median GDR with and without artifact removal was respectively 87.1 and 100. Median FA/h with and without artifact removal was 0.38 and 0. The difference in median FA/h was statistically significant ( p = 0.01). The difference in GDR was not statistically significant ( p = 0.50). 4 Discussion In this paper, we introduce an artifact removal strategy based on ICA–BSS to remove three types of biological artifacts from the neonatal EEG. These artifacts are ECG spikes, blood vessel pulsation and respiration activity. These hinder automated seizure detection and need to be removed. Our main goal was to reduce the number of false positive detections caused by these artifacts without lowering the sensitivity to detect the true seizures. This goal has been achieved. The number of FA/h has dropped significantly to 0 in this study and to 0.25 in an independent validation study ( Cherian et al., 2011 ). This implies that in a practical situation, a caregiver in an NICU needs to check the neonate only very few times based on a false alarm. The sensitivity to detect the neonatal seizures was not reduced by the preprocessing; in fact, it improved slightly but not significantly. The use of BSS for artifact removal as a preprocessing step is not novel. BSS methods became well known in EEG processing because of their superiority for removing eye movement artifacts ( Romero et al., 2008; Gao et al., 2010; Klemm et al., 2009 ), but also other artifacts were often well-separated. In this paper, we use different BSS implementations for different artifacts. This differentiation is based on theoretical considerations to match algorithmic properties with the characteristics of the artifact. A novel aspect of this paper is the way the artifactual signals are detected in an automatic way making use of additional polygraphic signals. The traditional automatic approach for detecting the artifactual component is by direct correlation with an EOG channel if such EOG reference channel is available ( Joyce et al., 2004; Mennes et al., 2010 ) or exploiting prior knowledge about topography or frequency characteristics ( Gao et al., 2010; De Vos et al., 2011; Ma et al., 2011 ). However, in reality there is no prior knowledge on which channel the ECG and respiration artifact will be strongest, so topographic information cannot be used for automatic identification. Moreover, the morphology is quite different between the artifact and the reference signal. We propose an effective way to use the correlation in common dynamics between extracted sources and reference signals for the automatic identification. This is a nice illustration how data fusion can make event detection more robust. Also in ( Mitra et al., 2009 ), correlation between the EEG traces and non-EEG traces is exploited to detect artifact segments. However, the authors did not exploit BSS methods so seizure activity covered by artifact cannot be detected. In this study, results on two groups of neonates were presented. One group had a lot of false alarms (Patient Nos. 1–9), where as in the second (Patient Nos. 10–13), no false detections were made, even without artifact preprocessing. Although we included more neonates having a lot of artifact-contaminated EEG, in daily practice, more neonates are expected to belong to the second group (without serious artifact contamination). However, as there are no clear physiological or technical indications of when the EEG of a certain neonate will thus be contaminated, it is important to take these possible artifacts into account. In an automated monitoring system, this BSS preprocessing step can also be set as an option that may be selected depending on the presence of biological artifacts that generate an excess of false alarms. For the results of the seizure detector with artifact preprocessing on a representative and independent sample of EEGs, we refer to Cherian et al. (2011) . In that study, also other performance values are presented (e.g., seizure burden) and the clinical impact of the seizure detector is discussed in detail. As can be seen in Table 2 , the seizure detector with the proposed artifact preprocessing still produces several false alarms. This is partly because there are more artifacts in the EEG than the ones we deal with in this chapter. The most important remaining artifacts are eye movements (nystagmus or pathologically fast eye movements) and tremor artifacts. Eye movements are relatively easy to recognize as they are always strongest on the frontal EEG channels. If the EOG is simultaneously measured, identification can be obtained by a simple correlation measure. Tremor artifacts can be detected with another strategy ( Matic et al., 2011 ). We also suggest developing methods to take these artifacts into account without removing them, as they contain clinical information about the neonate. In conclusion, we can say that the artifact removal as used in this paper is a beneficial addition to our automated seizure detector. For patients without artifacts, the artifact removal has no significant impact on the detector’s performance. For patients with the three major types of biological artifacts, the artifact removal significantly reduces the number of false positive detections, rendering the detection algorithms usable for continuous bed-side seizure monitoring. Acknowledgements Research supported by: Research Council KUL: GOA Ambiorics, GOA MaNet, CoE EF/05/006 Optimization in Engineering (OPTEC), PFV/10/002 (OPTEC), IDO 05/010 EEG–fMRI, IDO 08/013 Autism, IOF-KP06/11 FunCopt, several PhD/postdoc & fellow grants; • Flemish Government: FWO: Ph.D./postdoc grants, projects: FWO G.0302.07 (SVM), G.0341.07 (Data fusion), G.0427.10N (Integrated EEG–fMRI) research communities (ICCoS, ANMMM); IWT: TBM070713-Accelero, TBM070706-IOTA3, TBM080658-MRI (EEG–fMRI), Ph.D. Grants; IBBT • Belgian Federal Science Policy Office: IUAP P6/04 (DYSCO, ‘Dynamical systems, control and optimization’, 2007–2011); ESA PRODEX No. 90348 (sleep homeostasis) • EU: FAST (FP6-MC-RTN-035801), Neuromath (COST-BM0601). Appendix A Updates on the automated seizure detection Fig. A1 gives an overview of the detection strategy and implementation of the previously developed seizure detector ( Deburchgraeve et al., 2008 ). Briefly, we identified two seizure types, for each of which a separate detection algorithm was developed. The major difference between the two types is that the oscillatory type is a fluent, continuous seizure, whereas the spike train type consists of isolated spikes appearing on a background of lower voltage EEG. This means that the oscillatory type has a continuous kind of repetitiveness while the spike train type has a discontinuous kind. Another difference is that the oscillatory activity generally has lower frequency content than the high-frequency spikes. The updated components of the detection algorithm are shaded in grey in Fig. A1 . The most important update concerns a change in the segmentation strategy for the spike train type detection. For the oscillation detection, only the features extracted from the autocorrelation function have changed. Details on the other blocks can be found in our previous paper ( Deburchgraeve et al., 2008 ). A.1 Segmentation of the transients in the EEG The segmentation of the EEG transients is critical for the reliability of the seizure detection algorithm. This segmentation is performed on each channel of EEG, on a window of 5 s duration. There is an overlap of 4 s between subsequent windows under analysis. Fig. A2 shows a schematic overview of the updated algorithm. A.1.1 The non-linear energy operator (NLEO) We use the NLEO (Eq. 1) again to detect the local presence of high frequency activity. When applied to spike train type seizure EEG, the NLEO effectively amplifies the high-frequency spikes relative to the background EEG, facilitating their segmentation. A.1.2 Smoothing of the NLEO The NLEO calculates the energy of the signal based on only a few samples. However, the spikes of a neonatal spike train type seizure vary roughly between 50 and 500 ms. In order to match the sensitivity of the NLEO to the duration of the spikes, its output needs to be smoothed. However, it is not possible to find a single smoothing filter length that is adequate for both short (50 ms) and long (500 ms) spikes. This problem was solved by using a smoothing filter bank with 6 Moving Average (MA) filters with filter lengths of 2, 4, 8, 16, 32, and 64 samples. At a sampling frequency of 256 Hz, this leads to a set of filters in the order of ms up to 250 ms. The output of one filter is the input of the filter with next higher MA filter length. The output of the filter bank is the summation of the outputs of each filter. This leads to a smooth signal in which short as well as long spikes can easily be discriminated due to the varying lengths of the smoothing filters. Fig. A3C displays the smoothing effect on a spike train type seizure with long spikes (>500 ms). The arrows in Fig. A3B and A3C indicate that short peaks in the NLEO output are conserved by the smoothing: only variations of the NLEO output on a large time scale are smoothed out. This is exactly the behavior needed for the algorithm to be sensitive to both short and long spikes. A.1.3 Adaptive thresholding The goal of this step is to find a suitable threshold to discriminate between high and low energy. After thresholding, the parts of the signal with high energy are transformed to isolated segments with a certain position and length. The threshold must be at a level that detects the transients in the EEG without segmenting small, insignificant variations in the energy signal. For this purpose, segmentation is performed for a set of thresholds between 0 and 1 with a step size of 0.02. For each threshold, the number of segments above the threshold is counted. The threshold that leads to the maximum number of segments is kept as the definitive threshold. If several threshold levels lead to the same number of segments, the lowest one is taken. A.1.4 Zero elimination As a consequence of the previous segmentation step, a small dip in the smoothed energy signal may lead to a spike being segmented into two (or even more) isolated segments (see Fig. A3D . To correct this false segmentation, isolated low energy (zero) segments shorter than 100 ms are removed and combined with the high energy (one) segments surrounding it (see Fig. A3E ). This zero elimination may not be set too long, to avoid separate spikes close to each other being grouped into a single spike. A.2 Autocorrelation analysis The features that are extracted from the autocorrelation function in order to detect repetitive signals have also changed compared to our previous paper ( Deburchgraeve et al., 2008 ). In the updated version, three features are used to distinguish between repetitive and non-repetitive signals: - Regularity of the distances between the zero crossings ( Fig. A4A ), defined as ‘errorZeros’. - Regularity of the distances between the peaks ( Fig. A4B ), defined as ‘errorPeaks’. - Regularity of the normalized RMS values of the peaks which are delimited by the zero crossings ( Fig. A4C ), defined as ‘errorRMS’. On doing so, we make use of the fact that for an oscillatory signal, the phases of the autocorrelation function are regular. Hence, for oscillatory seizure activity, the above errors may be expected to be small. Regularity was measured by means of a pair wise comparison of all the distances or RMS values involved. For this purpose, each difference between an element indicated with a dark grey bar compared with that indicated by a light grey bars is expressed as a percentage of their difference in length or area. ( Fig. A4 ). For seizure detection, the thresholds on the features were defined as: - median([errorZeros,errorPeaks]) <7% and, - median(errorRMS) <10%. The comparisons for the zero crossings and the distances between the peaks can be grouped together, as both are distance measures. The comparisons for the RMS values are treated separately. All signals with properties below these thresholds are considered to be part of an oscillatory seizure. References Belouchrani et al., 1997 A. Belouchrani K. Abed-Meraim J.F. Cardoso E. Moulines A blind source separation technique using second order statistics IEEE Trans Signal Proc 45 1997 434 444 Cardoso and Souloumiac, 1993 J.F. Cardoso A. Souloumiac Blind beamforming for non-gaussian signals IEE Proc F 140 1993 362 370 Cherian et al., 2009 P.J. Cherian R.M. Swarte G.H. Visser Technical standards for recording and interpretation of neonatal electroencephalogram in clinical practice Ann Indian Acad Neurol 12 2009 58 70 Cherian et al., 2011 Cherian PJ, Deburchgraeve W, Swarte RM, De Vos M, Govaert P, Van Huffel S, et al., Validation of a new automated neonatal seizure detection system: a clinician’s perspective. Clin Neurophysiol 2011; doi: 10.1016/j.clinph.2011.01.043 . Choi et al., 2005 S. Choi A. Cichocki H. Park S. Lee Blind source separation and independent component analysis: a review Neural Inf Process Lett Rev 6 2005 1 57 Deburchgraeve et al., 2008 W. Deburchgraeve P.J. Cherian M. De Vos R.M. Swarte J.H. Blok G.H. Visser Automated neonatal seizure detection mimicking a human observer reading EEG Clin Neurophysiol 119 2008 2447 2454 De Clercq et al., 2006 W. De Clercq A. Vergult B. Vanrumste W. Van Paesschen S. Van Huffel Canonical Correlation analysis applied to remove muscle artifacts from the electroencephalogram IEEE Trans Biomed Eng 53 2006 2583 2587 De Vos et al., 2007 M. De Vos A. Vergult L. De Lathauwer W. De Clercq S. Van Huffel P. Dupont Canonical decomposition of ictal scalp EEG reliably detects the seizure onset zone Neuroimage 37 2007 844 854 De Vos et al., 2010 M. De Vos S. Ries K. Vanderperren B. Vanrumste F.-X. Alario S. Van Huffel Removal of muscle artifacts from EEG recordings of spoken language production Neuroinformatics 8 2010 135 150 De Vos et al., 2011 M. De Vos L. De Lathauwer S. Van Huffel Spatially Constrained ICA algorithms with an application in EEG processing Signal Process 91 2011 1963 1972 Gao et al., 2010 J.F. Gao Y. Yang P. Lin P. Wang C.X. Zheng Automatic removal of eye-movement and blink artifacts from EEG signals Brain Topogr 23 2010 105 114 Hellstrom-Westas et al., 1985 L. Hellstrom-Westas I. Rosen N.W. Swenningsen Silent seizures in sick infants in early life. Diagnosis by continuous cerebral function monitoring Acta Paediatr Scand 74 1985 741 748 Hyvärinen et al., 2001 A. Hyvärinen J. Karhunen E. Oja Independent component analysis 2001 John Wiley & Sons James and Hesse, 2004 C.J. James C.W. Hesse Independent component analysis for biomedical signals Physiol Meas 26 2004 15 39 Joyce et al., 2004 C.A. Joyce I.F. Gorodnitsky M. Kutas Automatic removal of eye movements and blink artefacts from EEG data using blind component separation Psychophysiology 41 2004 313 325 Kaiser, 1990 J.F. Kaiser On a simple algorithm to calculate the energy of a signal IEEE Int Conf Acoust Speech Signal Process (ICASSP) 1990 381 384 Klemm et al., 2009 M. Klemm J. Haueisen G. Ivanova Independent component analysis: comparison of algorithms for the investigation of surface electrical brain activity Med Biol Eng Computing 47 2009 413 423 Lanska and Lanska, 1996 M.J. Lanska D.J. Lanska Neonatal seizures in the United States: results of the National Hospital Discharge Survey, 1980–1991 Neuroepidemiology 15 1996 117 125 Ma et al., 2011 J. Ma S. Bayram P. Tao V. Svetnik High-throughput ocular artifact reduction in multichannel electroencephalography (EEG) using component subspace projection J Neurosci Methods 196 2011 131 140 Matic et al., 2011 Matic V, Cherian PJ, De Vos M, Visser G, Van Huffel S. Distinguishing tremor artifacts from neonatal epileptic seizures. Abstract, 3rd Dutch conference of biomedical engineering. Egmond aan Zee, Feb. 20–21, 2011. Mennes et al., 2010 M. Mennes H. Wouters B. Vanrumste L. Lagae P. Stiers Validation of ICA as a tool to remove eye movement artifacts from EEG/ERP Psychophysiology 47 2010 1142 1150 Mitra et al., 2009 J. Mitra J.R. Glover Y.P. Ktonas A.T. Kumar A. Mukherjee N.B. Karayiannis A multistage system for the automated detection of epileptic seizures in neonatal electroencephalography J Clin Neurophysiol 26 2009 218 226 Murray et al., 2008 D.M. Murray G.B. Boylan I. Ali C.A. Ryan B.P. Murphy S. Connolly Defining the gap between electrographic seizure burden, clinical expression and staff recognition of neonatal seizures Arch Dis Child Fetal Neonatal 93 2008 187 191 Nam et al., 2002 H. Nam T.G. Yim S.K. Han J.B. Oh S.K. Lee Independent component analysis of ictal EEG in medial temporal lobe epilepsy Epilepsia 43 2002 160 164 Onton et al., 2006 J. Onton M. Westerfield J. Townsend S. Makeig Imaging human EEG dynamics using independent component analysis Neurosci Biobehav Rev 30 2006 808 822 Romero et al., 2008 S. Romero M.A. Mañanas M.J. Barbanoj A comparative study of automatic techniques for ocular artifact reduction in spontaneous EEG signals based on clinical target variables: a simulation case Comput Biol Med 38 2008 348 360 Strobach et al., 1994 P. Strobach K. Abraham-Fuchs K. Harer Event-synchronous cancellation of the heart interference in biomedical signals IEEE Trans Biomed Eng 41 1994 343 350 Tang et al., 2004 A.C. Tang M.T. Sutherland C.J. Mckinney Validation of SOBI components from high-density EEG Neuroimage 25 2004 539 553 Temko et al., 2011a A. Temko E. Thomas W. Marnane G. Lightbody G. Boylan EEG-based neonatal seizure detection with Support Vector Machines Clin Neurophysiol 122 2011 464 473 Temko et al., 2011b A. Temko E. Thomas W. Marnane G. Lightbody G.B. Boylan Performance assessment for EEG-based neonatal seizure detectors Clin Neurophysiol 122 2011 474 482 Vanderperren et al., 2010 K. Vanderperren M. De Vos J.R. Ramautar N. Novitskiy M. Mennes S. Assecondi Removal of BCG artifacts from EEG recordings inside the MR scanner: a comparison of methodological and validation-related aspects NeuroImage 50 2010 920 934 Volpe, 2008 J.J. Volpe Neurology of the newborn 5th ed. 2008 Saunders WB Philadelphia Zarzoso and Comon, 2010 V. Zarzoso P. Comon Robust independent component analysis by iterative maximization of the kurtosis contrast with algebraic optimal step size IEEE Trans Neur Netw 21 2010 248 261