Fan Fiction

A method for the blind separation of sources for use as the first stage of a neonatal seizure detection system

Description
A method for the blind separation of sources for use as the first stage of a neonatal seizure detection system
Categories
Published
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
Share
Transcript
  A METHOD FOR THE BLIND SEPARATION OF SOURCES FOR USE AS THE FIRSTSTAGE OF A NEONATAL SEIZURE DETECTION SYSTEM S. Faul, L. Marnane, G. Lightbody University College Cork Dept. of Electrical and Electronic EngineeringCork, Ireland G. Boylan, S. Connolly Cork University HospitalDept. of Paediatrics and Child HealthCork, Ireland ABSTRACT A method is proposed for automatically choosing indepen-dent components (ICs) of interest from neonatal EEG data,with the aim of using them in further analysis to detectseizures. This procedure greatly reduces the amount of in-formation which needs to be processed in the seizure detec-tion system, and reduces the effect of noise and artefacts onits performance. The Fast ICA algorithm is used to generatethe ICs, and the complexity of each IC is examined to deter-mine those of interest. The Singular Value Fraction (SVF)measure is used to reduce the number of sources contain-ing artefacts chosen. In the best case, the 12 channel EEGused in these tests is reduced to 2 or 3 sources of interest. Inevery case, at least 3 sources were removed that consistedof noise. 1. INTRODUCTION Although systems for the detection of epileptic seizures inadults have been designed with varying degrees of success,detection of the neonatal seizure is a far more complex op-eration. Because of the ongoing development of the brain atthis age, seizure activity in neonates displays far more com-plex characteristics than in adults. Much of the analysis of neonatal seizures to date has concentrated on single-channelanalysis [2, 3]. However, it is clear from the methodologywhich EEG experts use to classify neonatal EEG that a reli-able neonatal seizure detection system will have to incorpo-rate multi-channel analysis.As a first stage to such a system it would be useful toreduce the amount of information which needs to be exam-ined in depth, and concentrate analysis on data where somedeviation from normal activity is evident. The EEG beingexamined is recorded at 200Hz in 12 channels, resulting in2400 samples per second. The availability of so much data,the computational load associated with the real-time appli-cation of detection algorithms becomes excessive leading tohardware solutions which are larger, more complicated andhence expensive. If elements of interest could be extractedfrom the EEG and analysed without having to analyse back-ground EEG activity, it would greatly decrease the compu-tational burden and simplify the seizure detection process.Independent Component Analysis (ICA) is a techniquefor separating an observed set of signals into a set of sta-tistically independent source signals, or independent com-ponents (ICs). Using this technique background activity,artefacts and seizure activity can be separated into differ-ent ICs. However, the major disadvantage with ICA is thatthe resulting ICs are not ordered in any way, and hence amethod is needed to extract the ICs of interest at the output.In this paper, it is proposed to use a combination of the complexity measure ( Ω ) developed by Roberts et al. [4]and the Singular Value Fraction (SVF) measure proposedby Kember and Fowler [5] to determine the ICs of interest. 2. METHODS2.1. Independent component analysis TheICAseparationprocessiscarriedoutwithoutpriorknowl-edge of the distribution of the sources, and is hence denotedBlind Source Separation (BSS). ICA can be seen as an evo-lution of Principle Component Analysis (PCA). However,ICA uses higher order statistics than PCA, and can find in-dependent sources in cases where PCA fails.There are many implementations of ICA techniques. Inthis paper the FastICA algorithm [6] is used. This approachis well documented and used widely in this field of research.It is straightforward to implement, fast and efficient.The main disadvantage to ICA is that, unlike PCA, theoutput ICs are not ordered in any context. Hence they mustbe examined further to extract the source(s) suited to theapplication. One method used to extract particular ICs isto use a reference signal which mimics the shape and tim-ing of the desired source. This approach is known as Con-strained ICA (cICA), and is used in artefact removal algo-rithms [7]. However, as the first stage of a proposed seizuredetection system for neonates, it is only necessary to isolateICs that show evidence of activity, seizure or artefact, which  will then be passed on for further tests. Also, using refer-ence signals to remove artefacts may inadvertently removeneonatal seizure activity as some of the reference signalsmay be highly correlated with seizure sources as well asartefact sources. 2.2. Embedding space decomposition To examine the ICs we follow the method proposed by Tak-ens [8] and perform an embedding space decomposition. Mdata points are selected from the IC and a trajectory matrix X  traj  of dimension  d E   × N   is then constructed, where  d E  is the embedding dimension and  N   =  M   − d E   + 1 .The rows of the trajectory matrix are made up of em-bedding vectors constructed by X  i  = [ x i − ( d E − 1)∆ ,x i − ( d E − 2)∆ ...x i ] T  ( d E   ≤ i ≤ M  ) (1)where  ∆  is the lag measured in number of data points.  ∆ and  d E   are chosen using a plot of   ∆  versus the SingularValue Fraction [5]. From analysis of these plots over a setof neonatal EEG it was calculated that  ∆ = 1  and d E   = 20 .The trajectory matrix is then composed by X  traj  = [ X  d E ,X  d E +1 ...X  M  ] T  (2) 2.3. Complexity analysis Performing singular value decomposition (SVD) on the tra- jectory matrix  X  traj  the singular values  σ 1 ...σ N   can befound. Using the methods of Roberts et al. [4] the entropyof the singular spectrum is defined by first normalising thesingular values such that ¯ σ j  =  σ j /  i σ i  (3)for  j  = 1 ...N  , and then defining the entropy H   = − N   i =1 ¯ σ i  log ¯ σ i  (4)The complexity of the data in each IC is measured by thenumber of states  Ω  and is defined as Ω = 2 H  (5)From previous work in the area of analysis of epilepticseizures in adults [9] it was shown that at epileptic seizureonset the number of states  Ω  decreased in ICs containingseizure activity, although no automatic means of extract-ing the ICs containing seizure activity has been developed.Therefore in this analysis  Ω  will be used to search for theICs of interest. 2.4. Singular value fraction Although the work carried out in complexity analysis showsthat  Ω  drops at seizure onset, it is clear that it also drops forartefacts. Although the aim at this point is not to separate allartefacts from the data, it would be an advantage to removethe more obvious artefacts in the process.From examination of singular values from sections of neonatal EEG, it is clear that there is a certain trend for thevalues obtained from non-seizure, seizure and artefact EEG.In [5] the SVF term is defined which gives the fractionalpower in the first k singular values. The SVF is defined as SVF  ( k ) = 1 − 1( d E   − k ) N  d E  i = k +1 σ 2 i  (6)The choice of k is discussed in [5] and either  k  = 1  or k  =  d A / 2  is suggested (where  d A  is the number of indicesfor which σ i  > δ  , some small noise threshold). In this study k  = 1  was used.The SVF shows a pronounced change in value in thepresence of artefacts, more so than the change in  Ω  and cantherefore be used to signify those ICs in which the artefactsappear. 2.5. Choosing/excluding ICs The ICs that are of interest for use in the seizure detectionprocess are, in general, the ICs with lower complexity,  Ω ,and little change in the SVF.Each IC in turn is windowed and the median value of   Ω is calculated over that window. ICs which contain noisewithout any significant information have much higher  Ω values and these are separated at this point by clusteringthe median values. The remaining ICs median  Ω  values arescaled from the IC with the minimum median  Ω , scaled to0, to the IC with the maximum median  Ω , scaled to 1.ThevarianceoftheSVFiscalculatedoneachwindowedIC (except those excluded by the clustering operation). ForICscontainingisolated, unwantedactivitytheSVFwillhavea largevariancecompared to those ICs thatcontain informa-tion of interest. The variance of the SVF is also scaled asdescribed above. The scaled values for  Ω  and the SVF arethen added together, giving a value close to 0 for ICs whichcontain seizure information and values close to 2 for thosewhich contain no traces of seizure activity. If the total foran IC is less than 1 it is selected as being ’of interest’.Itisobviousatthispointthatthisstagemustbedesignedto select too many ICs rather than too few (high sensitivity,low selectivity). If too few ICs are chosen then seizure in-formation could be lost and this would lead to a poor detec-tion rate for the system as a whole.  3. RESULTS A total of 4 hours of seizure data from 4 neonates was cho-sen for evaluation of this technique. All data was collectedfrom newborn babies with seizures in the neonatal inten-sive care units of Kings College Hospital in London, UKandCorkUniversityMaternityHospital, Ireland. ATelefac-tor Beehive video-EEG system or a Taugagreining NervusMonitor was used to record 12 channels of EEG using the10-20 system of electrode placement modified for neonates( F  4 − C  4 ,C  4 − P  4 ,P  4 − O 2 ,F  3 − C  3 ,C  3 − P  3 ,P  3 − O 1 ,T  4 − C  4 ,C  4 − C  z ,C  z − C  3 ,C  3 − T  3 ,T  4 − O 2 ,T  3 − O 1 ). A videorecording was made of each neonate for the duration of thestudy. A clinical neurophysiologist identified and classifiedall periods of seizure activity in each EEG recording.In all cases those ICs which could be seen to hold themajority of the seizure information were picked out suc-cessfully by the selection algorithm. In many cases an ICcontaining a low frequency near-sinusoidal signal was alsochosen by the algorithm. This signal is hypothesised, fromits frequency and morphology, to be a trace of the neonatesrespiration. Although it could be removed, the overall per-formance of the algorithm is not affected. In some casesICs secondary to the main information bearing IC that werealso deemed of interest were not selected by the algorithm.However, in all of these cases ICs with similar informationwere selected, and no loss in performance was suffered.In cases where only very few ICs contained informationof interest, there was a corresponding reduction in the num-ber of ICs selected. In the best of these cases the amountof data was cut from the initial 12 ICs down to 3 or 4, areduction of  ∼ 70% . Even in cases where the seizure activ-ity was evident across nearly all of the ICs, there were still3 or 4 ICs containing noise which could be excluded fromfurther analysis, hence still reducing the amount of data inthe worst cases by ∼ 30% . The algorithm was successful inrejecting ICs containing isolated bursts of activity withoutseizure information.Fig. 1 shows an example of 20 seconds of seizure EEGfrom a neonate. The seizure is evident on multiple chan-nels and there is a burst of unrelated activity at 2000 sam-ples. Fig. 2 shows the ICs calculated, and it is clear that theseizure activity has been sourced to one primary and onesecondary IC (marked by the arrows). ICs containing noisecan clearly be seen in ICs 9 through 12. It is also clear thatthe short burst of activity mentioned above affects the ’non-seizure’ ICs more than the others, as expected.Fig. 3 shows the  Ω  and SVF values calculated for eachIC (seizure ICs in bold). The  Ω  values for the primaryseizure IC are considerably less than the others, and the val-ues for the secondary IC are also low. The  Ω  values for the’noisy’ ICs are clearly separated from the others towardsthe top of the plot. The SVF values show the effect of theburst of activity that was seen in Figs. 1& 2 as large de-creases in the SVF values for the respective ICs. However,the two seizure ICs are not affected by the burst, and hencetheir variances are considerably lower. The ’noise’ ICs lieat the bottom of the plot with much lower SVF values. Ta-ble 4 shows the results for each IC including the score re-ceived from the sum of the two scaled measures (section2.5) and the resulting action. IC 6 had both the lowest me-dian number of states and the lowest variance in SVF andhence scored 0. IC 2 was the only other IC to score under 1. 4. CONCLUSIONS As a first stage to a seizure detection system for neonates,a method of data reduction is needed wherein no importantinformation is lost. This approach utilises ICA to obtainstatistically independent sources and a complexity measure Ω  and the SVF to choose the ICs of interest.Although studies have previously been carried out us-ing ICA to examine EEG, in most cases these use referencesignals to find artefacts or spikes in the EEG. In this studyno a priori information was assumed, making this method amore robust alternative.Aswellassimplybeingadatareductionprocess, theuseof the median  Ω  and the variance of the SVF allows the ex-clusion of ICs containing artefacts and noise, which shouldincrease the performance of the seizure detection system asa whole over a system which analyses raw EEG.Many routines were tested for choosing the appropriateICs. Originally a threshold approach was tested to choosetheICs, butthisonlygavesimplepositiveornegativeresultsfor each IC. Also the trends in  Ω  and the SVF shift from oneEEG sample to the next, making it difficult to choose robustthresholds. A ranking routine was also tested, but while itdid order the outputs by how likely they were to containinformation of interest, it gave no means by which to cut off the number of ICs being selected.The scaling routine provides a ranking of the ICs, butalso a value from which the number of ICs to be chosen canbe derived. The use of a threshold along with the scalingroutine means that the threshold shifts in relation to the in-put data, providing a consistent selection of the correct ICs.Before its implementation in a neonatal seizure detectionsystem, further tests are being carried out in this techniqueusing a larger data set. 5. REFERENCES [1] ClementPang, AdrianUpton, GlennShine, andMarkadKamath, “A comparison of algorithms for detection of spikes in the electroencephalogram,”  IEEE Transac-tions on Biomedical Engineering , vol. 50, no. 4, Apr.2003.  −1000100    f   4 −  c   4 −1000100   c   4 −  p   4 −50050   p   4 −  o   2 −1000100    f   3 −  c   3 −50050   c   3 −  p   3 −50050   p   3 −  o   1 −1000100    t   4 −  c   4 −1000100   c   4 −  c  z −50050   c  z −  c   3 −50050   c   3 −   t   3 −1000100    t   4 −  o   2 05001000150020002500300035004000−50050Samples (@200Hz)    t   3 −  o   1 Fig. 1 . Raw EEG −10010−10010−10010−10010−10010−505−10010−505−505−505−50505001000150020002500300035004000−505Samples (@200Hz)1 2 Fig. 2 . Independent Components 5101520253033.544.555.5Sections of EEG (4secs each)    N  u  m   b  e  r  o   f   S   t  a   t  e  s IC 1IC 2IC 3IC 4IC 5IC 6IC 7IC 8IC 9IC 10IC 11IC 12510152025300.20.30.40.50.60.70.8    S   V   F Sections of EEG (4secs each) Fig. 3 . Number of States ( Ω ) and the SVFIC No. Sum of Measures Result1 1.1724 Rejected2 0.8083 2nd Selected3 1.7266 Rejected4 1.4482 Rejected5 1.0912 Rejected6 0.0 1st Selected7 1.7187 Rejected8 1.5873 Rejected9 – Noise10 – Noise11 – Noise12 – Noise Fig. 4 . Result for each IC[2] Stephen Faul, Geraldine Boylan, Sean Connolly, LiamMarnane, and Gordon Lightbody, “Computer-aidedseizure detection in newborn infants,” in  2004 IEE IrishSignals and Systems Conference , Belfast, Northern Ire-land, June 2004, pp. 428–433.[3] B. Boashash, P. Barklem, and M. Keir, “Detectionof seizure signals in newborns,” in  1999 IEEE Inter-national Conference on Acousitcs, Speech and SignalProcessing , 1999, vol. 4, pp. 2351–2354.[4] Stephen J. Roberts, William Penny, and Iead Rezek,“Temporal and spatial complexity measures for eeg-based brain-computer interfacing,”  Medical and Bio-logical Engineering and Computing , vol. 37, no. 1, pp.93–99, 1998.[5] G. Kemberand A.C. Fowler, “Acorrelation function forchoosing time delays in phase portrait reconstructions,” Physics Letters A , vol. 179, pp. 72–80, 1993.[6] Aapo Hyvarinen and Erkki Oja, “A fast fixed-point al-gorithm for independent component analysis,”  NeuralComputation , vol. 9, pp. 1483–1492, 1997.[7] Christopher J. James and Oliver J. Gibson, “Tempo-rally constrained ica: an application to artifact rejectionin electromagnetic brain signal analysis,”  IEEE Trans-actions on Biomedical Engineering , vol. 50, no. 9, pp.1108–1116, Sept. 2003.[8] Floris Takens, “Detecting strange attractors in turbu-lence,”  Lecture Notes in Mathematics: Dynamical Sys-tems and Turbulence , vol. 898, pp. 366–381, 1981.[9] Christopher J. James and David Lowe, “Using indepen-dent component analysis and dynamical embedding toisolate seizure activity in the eeg,” in  Proceedings of the 22nd Annual EMBS International Conference , July2000.
Search
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