Detection of neonatal EEG seizure using multichannel matching pursuit

Detection of neonatal EEG seizure using multichannel matching pursuit
of 4
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  Detection of Neonatal EEG Seizure using Multichannel Matching Pursuit M. S. Khlif, M. Mesbah, B. Boashash, and P. Colditz Perinatal Research Centre, University of Queensland, Australia  Abstract   —  It is unusual for a newborn to have the classic “tonic-clonic” seizure experienced by adults and older children. Signs of seizure in newborns are either subtle or may become clinically silent. Therefore, the electroencephalogram (EEG) is becoming the most reliable tool for detecting neonatal seizure.   Being non-stationary and multicomponent, EEG signals are suitably analyzed using time-frequency (TF) based methods. In this paper, we present a seizure detection method using a new measure based on the matching pursuit (MP) decomposition of EEG data. Signals are represented in the TF domain where seizure structural characteristics are extracted to form a new coherent TF dictionary to be used in the MP decomposition. A new approach to set data-dependent thresholds, used in the seizure detection process, is proposed. To enhance the performance of the detector, the concept of areas of incidence is utilized to determine the geometrical correlation between EEG recording channels. 1.   INTRODUCTION Seizures are the most common and important sign of acute neonatal encephalopathy [1]. They are a major risk for death or subsequent neurological disability, and by themselves may contribute to an adverse neurodevelopment outcome. The incidence of seizures in infants born at term is 0.15 – 0.3%; the incidence is even higher in preterm infants, ranging from 5 – 15% [2]. About 50% of all survivors who have experienced neonatal seizure will develop long term complications like epilepsy and mental retardation [3]. Many clinicians have the assumption that neonatal seizures are epileptic in nature. Neonates, however, do exhibit a broad range of non-epileptic paroxysmal activities; making the distinction of epileptic from non-epileptic events a difficult task. Therefore, an automated system that would reliably detect neonatal seizure would be of significant clinical benefit in the neonatal intensive care unit. It is anticipated that such a system will remove the subjectivity and reduce errors associated with manual chart labeling, uncover subtleties in the EEG not otherwise detectable through visual inspection, and will enable continuous monitoring. Even though seizure EEG is nonstationary and can be multicomponent, many detection methods were designed using time or frequency domain-based signal analysis techniques [4]. Alternatively, more appropriate analysis can  be based on the joint TF domain. Careful investigation of TF representation of EEG data revealed significant differences existing in the structure and content of EEG signals between background and seizure states. Background EEG resembles noise in general and possesses much more complex structures compared to seizure EEG. The difference in structural complexity will be exploited to define a new structural complexity measure (SCM). We define the new SCM based on the matching pursuit (MP) decomposition of EEG data using a coherent TF dictionary. Based on SCM, we propose the design of the following four-step seizure detection process. First, seizure structures are extracted from the quadratic TF distribution (QTFD) of seizure EEG and a MP dictionary is constructed using waveforms having seizure-like structures. Next, the MP decomposition is performed and SCM vectors are determined for each EEG recording channel. Thirdly, SCM values are tested against a preset data-dependent threshold and outcomes are used to classify the EEG records. Finally, a technique, based on geometrical correlation among the individual channels, is used to enhance the performance of the seizure detector. 2.   SEIZURE STRUCTURE Quadratic TF distributions are popularly used to represent nonstationary signals in the joint TF domain. To extract seizure structures to be the basis for designing a coherent TF dictionary, a proper QTFD is chosen to visualize the EEG signals. Next, we present a brief overview of QTFD as well as a procedure for constructing a coherent TF dictionary. 2.1.   Quadratic TFD A QTFD reveals the number of components forming a signal, shows the energy concentration of each component, and efficiently spreads energy of white noise over the entire TF plane. TFDs differ mainly in their ability to suppress cross-terms and to resolve close auto-terms. On one end, the spectrogram is popular for being easy to interpret and for its efficient suppression of cross-terms; but at the cost of a reduced resolution. On the other end, the Wigner-Ville distribution (WVD) is viewed as the best TFD for linear FM signals. But despite its many desirable properties, WVD becomes difficult to interpret in the case of multicomponent signals or in the presence of noise. The class of reduced-interference distributions (RID), obtained by smoothing the WVD, attempts a compromise  between the efficient cross-term suppression of the spectrogram and the ideal energy concentration of the WVD. The general form of RID is given by [5] 2 (,)(,)()*()22  j f  z   t f G t u z u z u e dud  π τ   τ τ   ρ τ τ   − = − + − ∫∫   (1) where  z*  stands for the complex conjugate of  z;  the analytical associate of a real signal  x(t) . The smoothing 30th Annual International IEEE EMBS ConferenceVancouver, British Columbia, Canada, August 20-24, 2008 978-1-4244-1815-2/08/$25.00 ©2008 IEEE.907  kernel G(t, τ   )  defines a particular TFD. For EEG signal representation, we found the modified B distribution (MBD) to be most appropriate. MBD kernel is given by 22 cosh(,),01cosh t G t d   β  β  τ β ξ ξ  −− = < < ∫   (2)    β    controls the trade-off between TF resolution and cross-term suppression. 2.2.   Coherent dictionary design Earlier analysis of background and seizure EEG by the  present authors concluded that seizure EEG can generally  be characterized by linear frequency modulated (LFM) or  piece-wise LFM patterns [6].   For a seizure signal 8 to 12 seconds long, the signal can be adequately modeled as a LFM with a frequency slope in the range of [-0.05,+0.05] Hz/sec. On the contrary, background EEG reveals no specific patterns. Therefore, it seems logical that the design of a coherent decomposition dictionary should include a set of LFM atoms; which are given by 2 ()exp ()2()2(1)  LFM   f ii n n j n N  φ  ξ ξ π ξ θ  =    −+ +     −       (3) where ()()1  LFM LFM  φ φ  ℜ = ℑ = ,  N    is the discrete signal sample size, θ    is an initial phase , and   (,)[0,0.5] i f  ξ ξ   ∈   are respectively initial and final normalized frequencies. This dictionary will be referred to as the “LFM TF dictionary”. 3.   MP ALGORITHM Matching pursuit decomposition with TF dictionaries was srcinally proposed by Mallat and Zhang [7] to sub optimally represent signals in an overcomplete dictionary. The MP algorithm was designed to decompose any signal into a linear expansion of elementary functions. In the formulation of matching pursuit, the word atom [ () t  γ   φ  ]   refers to a single waveform used to represent a signal of interest. A collection of atoms forms the dictionary [ { } γ   φ  Φ = ]. When Φ  spans the Hilbert space [ 2 ()  L R ] of complex valued functions, it is said to be complete. If the dictionary contains more atoms than are needed to span the Hilbert space, the dictionary is said to be over-complete or redundant . A TF dictionary is composed of scaled ()  s , translated () u , and modulated () ξ   window function () t  φ  . In [7], the TF dictionary atoms were defined  by 21 () ()  j t t u s s t   e γ   πξ  φ   φ   − =  (4) where { } ,,  s u γ ξ  = and ()1 t  φ   = . The popular Gabor TF dictionary is characterized by the transformations of a fundamental Gaussian window function, 2 1/4 ()2  t  t   e  π  φ   − =  through a set of indexes γ   . To make it redundant, the Gabor dictionary was combined with delta and complex Fourier functions; referred to here as “modified Gabor”. In MP decomposition, the atoms which correlate the most to the signal being decomposed are iteratively and optimally selected from the redundant dictionary. In terms of selected atoms, the decomposed signal is represented by -1-100 , m mm i mi i i ii i  x R x R x R x γ γ γ γ   α φ φ φ  = = = + = + ∑ ∑   (5) where the coefficient i γ   α   is the inner product of the selected atom i γ   φ   and the residual signal i  R x . m  R x is the residual signal after m  iterations. 4.   MP APPLICATION TO EEG 4.1.   Decomposition criteria: MP is an iterative process and thus requires a stopping criterion. The first criterion defined in [7] was based on limiting decomposition to coherent structures only. This meant stopping the process after the first m  atoms that have a higher than average correlation with the signal residue. Later, Jouny et. al. [8] imposed a second criterion based on energy of the last atom added. The new criterion guarantees that all signals are decomposed to the same “energy resolution”; even when signals have different energy levels. In this work, we opted for the frequently used criterion  based on percentage of total signal energy. As we are not seeking to fully explain the decomposed signal, this last criterion preserves an acceptable level to the discriminating capabilities of the EEG classifier. 4.2.   SCM definition Figure 1.    Number of atoms ( m ) needed for the decomposition of a single channel EEG using Gabor and LFM dictionaries. In [8, 9], it was demonstrated that the minimum number ( m ) of atoms required to decompose a signal to a desired criterion, is related to the complexity of that signal. In  principal, m  can then constitute a reliable measure to monitor changes in signal characteristics. In this work, we report similar findings and we choose to use m  to monitor changes between background and seizure states. From Fig.1, it is clear that the MP decomposition of seizure 908  epochs (identified in the middle plot) required fewer atoms   compared to background epochs. Further, Fig.1 showcases a structural complexity measure with superior discriminatory capability when based on the new LFM TF dictionary. This superiority is manifested in more compact decompositions during seizure periods and in higher m  values during periods of inactivity. Up to 71% reduction in m  has been achieved when seizure epochs were decomposed using the coherent LFM dictionary. Finally, normalizing the measure m  by the signal energy defines the new structural complexity measure as / Signal  SCM m E  =   (6)   4.3.   Threshold setting Although the trend of m  is consistent (fewer atoms for seizure) from one patient to the other, the mean level of m  is quite different between patients. It is therefore mandatory to define a data-dependent detection threshold if automated EEG classification is desired. Preliminary investigation has revealed that such a threshold could be set using { } ( ) { } ( ) min3,4,3,4min33,44 T median SCM C C P P Monopolar T median SCM C P C P Bipolar    = →    = − − →     (7) EEG data is available from a total of 14 monopolar channels, all referenced to a common remote location. 20  bipolar channels are formed by computing the difference  between pairs of monopolar channels (see Fig.2). The channels in (7) may be the most conveniently located to register the presence of a seizure. Finally, EEG classification is accomplished by testing SCM   against T.   SCM T seizure < →   (8) 5.   MULTICHANNEL SEIZURE DETECTION It is anticipated that an automated standalone seizure detection system is capable of functioning without a priori knowledge of seizure location or of the channels to be  probed for seizure presence. To blindly specify one or few individual channels for analysis would constitute an unreliable detection method. It is therefore imperative to systematically investigate and use all available EEG data (be it monopolar or bipolar). To design an unaided detection system, we make use of the geometrical mapping of the multichannel data. Referring to Fig.2, we define the locations of monopolar channels as areas of seizure incidence. Respectively, each area is either surrounded by or linked to a certain number of monopolar or bipolar channels; referred to as channels of adjacency. Regardless of which recording method is used, a seizure detection at a given area of incidence (e.g. “C4”) is said to be geometrically correlated when the surrounding monopolar channels (e.g. “2,6,10,13”) or the linked bipolar channels (e.g. “7,8,13,14) all record seizure detections. Monopolar: 1 F32 F43 C34 C45 P36 P47 O18 O29 T310 T411 T512 T613 Cz14 PzBipolar channels: 1:F4-T4,2:T4-T6,3:T6-O2,4:F3-T3,5:T3-T5,6:T5-O17:F4-C4,8:C4-P4,9:P4-O2,10:F3-C3,11:C3-P3,12:P3-O1,13:T4-C414:C4-CZ,15:CZ-C3,16:C3-T3,17:T6-P4,18:P4-PZ,19:PZ-P3,20:P3-T5 Monopol    ar: 1 F32 F43 C34 C45 P36 P47 O18 O29 T310 T411 T512 T613 Cz14 PzBipolar channels: 1:F4-T4,2:T4-T6,3:T6-O2,4:F3-T3,5:T3-T5,6:T5-O17:F4-C4,8:C4-P4,9:P4-O2,10:F3-C3,11:C3-P3,12:P3-O1,13:T4-C414:C4-CZ,15:CZ-C3,16:C3-T3,17:T6-P4,18:P4-PZ,19:PZ-P3,20:P3-T5   Figure 2.   Areas of incidence and channels of adjacency. Once EEG data from individual channels is classified using (8), geometrical correlation is used to obtain a multi-channel detection vector. The details about implementing this procedure are found in our earlier work [10]. 6.   RESULTS AND DISCUSSION Six EEG records were used to test the proposed seizure detector. Data acquisition and labeling were performed by a neurologist at the Royal Children’s Hospital, Brisbane, Australia. Data was acquired at a sampling rate of 256 Hz, then was down sampled to 32 Hz to reduce computation time. Since most neonatal EEG power is found in frequencies in the range of [0.5,4] Hz, data was bandpass filtered to this frequency range. Prerequisite to EEG classification, a modified Gabor and a coherent LFM TF dictionaries were both constructed. The MBD (  β   = 0.01) was used to represent seizure EEG in the TF domain. Then, the characteristics of the LFM atoms forming the coherent LFM dictionary, were extracted. The classification process consisted of: a) MP decomposition of 8 second-long sliding epochs (  N   = 256 samples) without any overlapping at this stage, b) implementing (6) to determine SCM   vectors, and c) applying (8) to form seizure detection masks for individual EEG channels. The MP decomposition criterion was set to 10% of signal energy. A detection mask is obtained by assigning “1s” to epochs identified in (8) as being seizure, “0s” are assigned otherwise. Detection masks were compared to neurologist labeling and detector performance was evaluated. To improve on performance, the classification outcome of a given epoch was not accepted unless it occurred in a succession of three outcomes of the same type. The  previous restriction practically made the detector more sensitive to seizure with added robustness against artefacts. Performance of the seizure detector is evaluated in terms of true positive and false positive rates  [TPR=TP/(TP+FN); FPR=FP/(TN+FP)] , where TP  : true  positive, TN  :   true negative,  FP  : false positive, and  FN  : false negative [10]. S=(TP+FN)  defines the total number of seizure epochs as marked by the neurologist, while  NS=(TN+FP)  is the total number of background epochs. The word “positive” refers to an epoch classified as seizure. Appropriateness of the detection threshold (7) is tested  by comparing T   to an optimum threshold determined by maximizing the following measure 909    ( ) ( ) ( ) 31/3/4,/ TPTNnnSNSSNS  α σ σ  σ   = + + += + =   (9) Table I provides a summary of this comparison proving that (7) can be used to reasonably approximate optimum detection thresholds obtained by maximizing (9). The measure in (9) attempts to find optimum values for TPR  and  FPR with the aim of maximizing TPR  and minimizing  FPR . Table I : Thresholds used in the detection process as compared to optimum thresholds. 1 2 3 4 5 6Optimum 1.00 0.20 2.50 3.40 5.90 1.40Process 0.75 0.30 2.60 4.10 6.00 1.40Optimum 0.60 0.30 5.80 3.80 4.70 2.00Process 1.20 0.40 3.80 5.00 5.00 1.90 Neonatal Case No.Monopolar EEGBipolar EEG   Table II shows a summary of the overall performance of the seizure detector based on the MP decomposition of six EEG records given different process conditions. Process  parameter variations include nature of the decomposition dictionary, type of EEG data (monopolar vs. bipolar), classifier type (single channel vs. multichannel), and type of threshold used (process vs. optimum). Performance data  based on optimum thresholds, included here only for reference, is not practical as it requires a priori labeling of the data. Table II : Performance summary for different process conditions. Dictionary Modified Gabor  Classifier  MultichannelSingle Channels Threshold Process ProcessOptimumProcess TPR (%) 86 88 88 87FPR (%) 7 5 4 9TPR (%) 88 90 91 87FPR (%) 4 5 2 9 Monopolar Bipolar  EEG Data Type Coherent LFM Multicahnnel   The performance figures shown above strongly indicate that SCM   correlates well with seizure dynamics and that the  proposed MP decomposition can effectively be used for neonatal seizure detection. Proposed methods repeatedly  produced detections masks very much similar to neurologist labeling except for some start and stop timing differences. Since labeling of EEG is largely a subjective matter [10] and can be affected by human inaccuracies,  performance figures in Table II may be well within the confidence levels of such labeling. Table II reiterates the fact that MP decomposition is dependent on the nature of the decomposition dictionary, and more importantly that a better seizure detection can be achieved with the proposed coherent LFM dictionary. The numbers also indicate that the exploitation of geometrical correlation between EEG channels is effective in reducing the rate of false detections (by 4%). Further, the method of multichannel detection lays itself to the implementation of an online process. Other conclusions drawn from Table II include the comparable performance between detections  based on optimum thresholds and data-dependent thresholds obtained using (7); especially in the case of monopolar data. The method of (7), still needed to be validated using a larger patient database, enables the design of automated detection despite large fluctuations in SCM   levels between patients. Table II also indicates a higher  performance in favor of bipolar EEG. This result may not  be sustainable since neurologist labeling was based on  bipolar data. 7.   CONCLUSION In this paper, we have defined a new structural complexity measure ( SCM  ) and used it to design a neonatal seizure detector. SCM   is based on the MP decomposition of EEG and is strongly dependent on the nature of decomposition dictionary used. Proper QTFD was used to represent and extract the characteristics of TF atoms that were used to construct a coherent LFM TF dictionary. As expected, the new dictionary was more coherent with seizure EEG dynamics and thus provided more efficient seizure detection. Automated seizure detection was only accomplished after the introduction of a new approach for setting the data-dependent thresholds being used in the detection process. The method efficiently computed thresholds that resulted in comparable performance when compared to optimum thresholds. A new measure ( α ) was introduced to help determine optimum thresholds. Finally, in exploiting the concept of areas of incidence to identify correlation between EEG channels, we founded a method to enhance the performance of the proposed seizure detector; resulting in 4% reduction in overall  FPR . REFERENCES   [1] R. R. Clancy, "Summary proceedings from the neurology group on neonatal seizures,"  Pediatrics, vol. 117, pp. S23-S27, Mar 2006. [2] J. J. Volpe, "Neonatal seizures," in  Neurology of the Newborn vol. 4th Edition, (ed JJ Volpe) Philadelphia: W.B. Saunders, 2000, pp. 129-159. [3] M. Wong, "Neonatal Seizures." vol. 2007: Epilepsy Center at St. Louis Children's Hospital, 2007. [4] S. Faul, G. Boylan, S. Connolly, L. Marnane, and G. Lightbody, "An evaluation of automated neonatal seizure detection methods," Clinical   Neurophysiology, vol. 116, pp. 1533-1541, Jul 2005. [5] B. Boashash, "Time-Frequency Signal Analysis and Processing: A Comprehensive Reference ", 1 ed Oxford, UK: Elsevier, 2003, pp. Chapters: 1-3. [6] B. Boashash and M. Mesbah, "A time-frequency approach for  newborn seizure detection "  IEEE Engineering in Medicine and   Biology Magazine, vol. 20, pp. 54-64, 2001. [7] S. G. Mallat and Z. F. Zhang, "Matching Pursuit with Time-Frequency Dictionaries,"  Ieee Transactions on Signal Processing, vol. 41, pp. 3397-3415, Dec 1993. [8] C. C. Jouny, P. J. Franaszczuk, and G. K. Bergey, "Characterization of epileptic seizure dynamics using Gabor atom density," Clinical  Neurophysiology, vol. 114, pp. 426-437, 2003. [9] L. Rankine, M. Mesbah, and B. Boashash, "A matching pursuit-based signal complexity measure for the analysis of newborn EEG," Med Bio Eng Comput, vol. 45, pp. 251-260, 2007. [10] M. S. Khlif, M. Mesbah, B. Boashash, and P. Colditz, "Multichannel-Based Newborn EEG Seizure Detection using Time-Frequency Matched Filter," in Engineering in Medicine and Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the IEEE, 2007, pp. 1265-1268.   910
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks