Abstract Electroencephalogram processing using neural networks

Electroencephalogram processing using neural networksClaude Robert a,b,*,Jean-Franc ¸ois Gaudy b ,Aime´Limoge a aLaboratoire d’E´lectrophysiologie,Universite ´Paris 5-Rene ´Descartes,1rue Maurice Arnoux,92120Montrouge,France bLaboratoire d’Anatomie Fonctionnelle,Universite´Paris 5-Rene ´Descartes,1rue Maurice Arnoux,92120Montrouge,France Accepted 23January 2002AbstractThe electroencephalogram (EEG),a highly complex signal,is one of the most common sources of information used to study brain functionand neurological disorders.More than 100current neural network applications dedicated to EEG processing are presented.Works are categorized according to their objective (sleep analysis,monitoring anesthesia depth,brain-computer interface,EEG artifact detection,EEG source-based localization,etc.).Each application involves a specific approach (long-term analysis or short-term EEG segment analysis,real-time or time delayed processing,single or multiple EEG-channel analysis,etc.),for which neural networks were generally successful.The promising performances observed are demonstrative of the efficiency and efficacy of systems developed.This review can aid researchers,clinicians and implementors to understand up-to-date interest in neural network tools for EEG processing.The extended bibliography provides a database to assist in possible new concepts and idea development.q 2002Elsevier Science Ireland Ltd.All rights reserved.Keywords :Electroencephalogram;Neural network;Review1.IntroductionSince the pioneer work of Berger (1929),the electroen-cephalogram (EEG)has become a proven source of infor-mation for clinicians and researchers.Early on,EEG analysis was restricted to visual inspection of EEG records.The development of electrical devices combined with the Fast Fourier Transform algorithm,oriented EEG analysis in terms of spectral components.During the 1970s,the intro-duction of micro-computer technology in medicine and biology revolutionized approaches to EEG processing.New methods of signal and data analysis as well as increas-ing powers of computing,provide improved tools to record and analyze the EEG.Among new computing tools,neural networks have been successfully employed to process EEG signals in order to study brain mechanisms by using multi-ple approaches.This paper presents a spectrum of neural network applications in EEG processing.An extended bibliography is included.2.Artificial neural networksArtificial neural networks (ANNs)are computational tools utilizing a combination of many elementary processing units (cell).Each unit is connected to a number of network units toprocess information by transfer function.The relationship between the input and the output determine network beha-vior.Contrary to conventional computing methods,ANNs are ‘trained’to produce desired the input-output relationship.During the training (learning)phase,examples of data are presented to the network and,using a learning algorithm,the parameters are tuned to adjust network behavior.According to available knowledge of the problem,and the objective of the operator,the learning procedure employed can be either ‘supervised’,‘unsupervised’,or both.The supervised learn-ing procedure is performed with pairs of known input-output patterns while unsupervised learning consists of presenting training examples to the network input and the network orga-nizes itself progressively to reach maximal separation between the naturally occurring classes of examples.The principal applications of ANNs have been in the area of pattern recognition.The pattern is turned into a feature vector used as ANN input,and the output is interpreted as identify-ing the input to be a member of one of a number of classes of possible inputs.An important quality of neural networks (referred to as generalization)is that when they are correctly trained,neural networks can appropriately process data that have not been used for training.The most popular neural networks used are multilayer perceptrons which are generally supervised-trained with the error back-propagation algorithm (Rumelhart et al.,1986).One major property of these networks is their ability to find non-linear surfaces separating the underlayingClinical Neurophysiology 113(2002)694–701CLINPH 2001553*Corresponding author.Tel.:133-1-5807-6798;fax:133-1-5807-6899.E-mail address:claude.robert@odontologie.univ-paris5.fr (C.Robert).patterns which is generally considered as an improvement on conventional methods.Radial basis function(RBF)network is a particular class of multilayer networks(Poggio and Girosi,1989)in which learning occurs usually in two stages,learning in the hidden layer(usually by an unsupervised bottom-up self-organizing method such as k-means clustering)followed by the output layer(a top-down supervised method such as least squares estimation).RBF networks present two important advan-tages:finding the input to output map using local approx-imators,and rapid learning requiring fewer examples.An other popular class of networks is the self-organizing map,or Kohonen-network(Kohonen,1988).A Kohonen network consists of two fully connected-unit layers.The output layer is generally ordered in a low-dimensional framework(a one-dimensional array or a two-dimensional matrix)of units.The objective is to build a map where units of an area are activated when inputs with similar character-istics are presented.Among the other popular networks are adaptive reso-nance theory(ART)networks and their derivates(ART1, ART2,fuzzy ART,etc.)(Carpenter and Grossberg,1992) and Hopfield models(Hopfield,1982).One quality of neural networks is that they can be consid-ered as non-linear statistical method.Nevertheless,a large amount of data are required to overcome the existing non-linearities in the data structure.As this condition is not always fulfilled in electroencephalography,it restricts the applications of neural networks in many cases.Further developments and details concerning neural networks theory and implementation are available in the literature(Kohonen,1988;Freeman and Skapura,1991; Hertz et al.,1991;Zurada,1992),and in medicine and biol-ogy reviews(Miller et al.,1992;Algaver et al.,1994;Sabba-tini,1996).3.EEG-artifact processingDetection and rejection of artifacts in the EEG is a diffi-cult and unpleasant task for which few neural network systems have been developed(Sadasivan and Dutt,1994; Wu et al.,1994a;Clochon et al.,1992;Durka et al.,1996; Riddington et al.,1996;Herrmann,1997;Saastamoinen et al.,1998;Bogacz et al.,1999;Vasko et al.,2000).The main result is the heterogeneous accuracies of the systems devel-oped.The treatment of EEG artifacts will probably become a primordial problem with increasing automation of signal processing.Though this unavoidable task cumulates diffi-culties,early positive results are encouraging signs for future applications.4.EEG data compressionA need for EEG compression is justified by the need for high accuracy measurements,long-term measurements,transmission of data in telemedicine or the lack of special equipment for storing voluminous datafiles.Nevertheless, the complexity of the EEG waveforms is the reason for few attempts at EEG compression.Preliminary studies invol-ving neural networks have been performed(Bargiotti et al.,1993;Battiti et al.,1995;Krajca et al.,1997).When high compression rates(from50to80%)were obtained (Battiti et al.,1995),further investigation was necessary to confirm the plausibility of neural network EEG compres-sion.5.EEG-based source localizationOne method useful in understanding brain dynamics and clinical diagnosis of brain abnormalities is EEG based source localization which involves localization of intracra-nial sources using EEGs by means of optimization techni-ques such as iterative methods.Once a neural network is trained,it no longer requires iterations,and is able to capture non-linear dynamics of the source localization problem, while maintaining the noise robustness essential to EEG analysis.The error position in source localization was less than5%for all investigations(Abeyratne et al.,1991,2001; Yuasa et al.,1998;Zhang et al.,1998;Van Hoey et al., 2000;Sun and Sclabassi,2000;Tun et al.,2000).While iterative methods(Levenberg-Marquardt algorithm)some-time provide better results than neural networks in the noise free situation,neural network methods perform better at low signal to noise ratios which is more realistic(Tun et al., 2000).Work is expected to continue in this promisingfield.6.Basic EEG processing(systems tested on healthy subjects,not patients)This paragraph includes applications in which investiga-tors focused on extracting information from EEG records without testing the system on patients.Some of the various ANN-based systems developed were aimed at detecting specific graphoelements:K-complex waves(Bankman et al.,1992),identification of EEG spikes or sleep spindles (Bankman et al.,1992;Reddy and Korrai,1992;Huuponen et al.,2000),or recognition of topographic patterns of EEG spectra(Emiliani and Frietman,1994;Joutsiniemi et al., 1995).Performances(correct recognition)of these systems varied from69(Emiliani and Frietman,1994)to89% (Bankman et al.,1992;Joutsiniemi et al.,1995)with a low rate(,8%)of false positive(Bankman et al.,1992; Huuponen et al.,2000).In another approach,the develop-ment of ANN-based tools to analyze evoked potentials was the challenge.When investigations were performed on simulated data(Uncini et al.,1990;Dumitras et al.,1994), most of the ANN-based systems were developed using data from evoked potentials recorded on healthy subjects.Some were concerned with visual evoked potential analysis(Van der Kouwe and Cilliers,1995;Fung et al.,1996;Laskaris etC.Robert et al./Clinical Neurophysiology113(2002)694–701695al.,1997;Fung et al.,1999;Leistritz et al.,1999a;Hoffmann et al.,2001),others were designed to analyze auditory evoked potentials(Bruha and Madhavan,1989;Bruha et al.,1990;Fung et al.,1999)or study somatosensory evoked potentials(Liberati,1991).Other systems were developed to analysis EEG records during cognitive processes on healthy subjects(Flexer et al.,1995;Lange et al.,2000). Though the performances of these systems are inhomo-geneous both intrinsically and in their presentation,a global positive appreciation emerges.Most of these systems need further investigation before operational applicability can be integrated in a clinical setting.7.Sleep studiesAmong the different parameters employed in sleep studies to establish hypnograms,the information carried by the EEG is of prime importance in determining the state of vigilance of the subject under consideration.The tedious and time consuming task required to build hypno-grams has generated development of new tools for clinicians and researchers.As the construction of hypnograms consists in associating thousands of epochs,to an appropriate state of vigilance,investigators have naturally made use of neural networks to enable correct processing of new data(general-ization).These systems studied sleep either in cat(Mamelak et al.,1991),rat(Robert et al.,1996)or human(Principe and Tome,1989;Pfurtscheller et al.,1992a;Roberts and Taras-senko,1992;Schaltenbrand et al.,1993;Gro¨zinger et al., 1995;Pardey et al.,1996;Sykacek et al.,1997;Baumgart-Schmitt et al.,1997;Stewart et al.,1999).Performance of these systems was inversely proportioned to the number of classes discriminated:the system trained to discriminate3 states in the rat(wake,slow wave sleep and paradoxical sleep)presented high global accuracy(.95%)(Robert et al.,1996)while global accuracy ranging from61to80% was observed in the system to discriminate7classes(wake, movement,sleep stage1,sleep stage2,sleep stage3/4, paradoxical sleep and artifacts)in babies(Pfurtscheller et al.,1992a).Though they do not provide optimal results, these systems have been successfully used in the clinical environment(Roberts and Tarassenko,1995;Gro¨zinger et al.,1995;Schaltenbrand et al.,1996)and in experimental conditions(Mamelak et al.,1991;Robert et al.,1999). These various applications demonstrate pertinency of neural networks in developing sleep staging systems.Readers interested in these applications are invited to read a synthetic publication(Robert et al.,1998).8.Monitoring depth of anesthesia or patients under sedation in the intensive care unitMonitoring patients in intensive care units or patients under anesthesia is anotherfield in which the EEG can provide clinical information of the state of the central nervous system.Recently,several EEG processing neural network-based systems have been proposed for monitoring anesthesia depth(Shuter et al.,1994;Krkic et al.,1996; Sharma and Roy,1997;Nayak and Roy,1998;Huang et al.,1999;Muthuswamy and Roy,1999;Zhang and Roy, 1999;Allen and Smith,2001;Zhang et al.,2001).Most of these investigations were performed at the Albany Medical College(NY,USA)on dogs under propofol(Huang et al., 1999;Zhang and Roy,2001),isoflurane(Nayak and Roy, 1998;Muthuswamy and Roy,1999;Zhang and Roy,1999) or halothane(Sharma and Roy,1997)anesthesia.Other investigations involving neural network-based systems were performed on humans(Allen and Smith,2001; Zhang et al.,2001).High degrees of performance(about 90%accuracy)characterize most of these studies(Krkic et al.,1996;Sharma and Roy,1997;Nayak and Roy, 1998;Huang et al.,1999;Muthuswamy and Roy,1999; Zhang and Roy,1999;Zhang et al.,2001).Other ANN-systems were developed to monitor sedation level of ICU critically ill patients(Veselis et al.,1991)or to detect EEG burst-suppression in patients under sedation(Leistritz et al., 1999b).Promising results encourage further investigation.9.Monitoring alterness and vigilanceAnother important application of EEG processing is the study of the time course of alterness and vigilance of opera-tors who perform monotonous but attention demanding tasks(air traffic controllers,lorry drivers,etc.).The objec-tive is to avoid potential accidents generated by decreased vigilance or cognitive impairment associated with intoxica-tion or fatigue using a real-time system which can continu-ously monitor vigilance,thereby preventing accidents caused by attention deficit.Using EEG recordings,several investigators developed neural network-based systems to assess the vigilance level of the subject under record(Belenky et al.,1994;Gulati et al.,1995;Koska et al.,1996;Makeig et al.,1996;Jouny, 1997;Jung et al.,1997;Kohlmorgen et al.,1997;Roberts et al.,2000).Most of these investigations are objective-oriented:assessment of alterness in the sleep-deprived subject(Belenky et al.,1994)or air-traffic controllers (Gulati et al.,1995),study of vigilance levels of subjects with chronic occupational chemical stress(Tuulik et al., 1997),detecting transition between different vigilance states (Jouny,1997;Kohlmorgen et al.,1997),or states of transi-ent cognitive impairments induced by mild acute intoxica-tion or fatigue(Gevins and Smith,1999).While some studies were performed on only one subject(Gulati et al., 1995;Kohlmorgen et al.,1997),optimistic conclusions encourage continued study.10.EpilepsyAnother domain of ANN-based system testing is theC.Robert et al./Clinical Neurophysiology113(2002)694–701 696study of epileptic phenomenon in the EEG to provide clin-ical information diagnosing,monitoring and managing related neurological disorders.The objective of many inves-tigators was to develop a system capable of detecting(Elo et al.,1992;Gabor et al.,1996;Pradhan et al.,1996;Webber et al.,1996;Varsta et al.,1997;Gabor,1998;Walczack and Nowack,2001)or predicting(Petrosian et al.,2000)seizure in the EEG.Other system goals were to detect epileptiform transient waveforms such as spikes and sharp waves typi-cally observed between seizures(Eberhart et al.,1989; Ozdamar et al.,1991;Gabor and Seyal,1992;Webber et al.,1994;Kalayaci and Ozdamar,1995;Jando et al.,1993; Park et al.,1998;Tarassenko et al.,1998;James et al.,1999; Kobayashi et al.,1999;Ko and Chung,2000).Among these investigations,some must be considered preliminary as EEG records used were from only a small number of patients(,10)(Elo et al.,1992;Pradhan et al.,1996;Varsta et al.,1997;Tarassenko et al.,1998;Kobayashi et al.,1999; Petrosian et al.,2000).Others were performed on a larger database(.20patients)(Ozdamar et al.,1992;Jando et al., 1993;Gabor et al.,1996;Webber et al.,1996;James et al., 1999;Ko and Chung,2000;Walczack and Nowack,2001) and were included in a continuum process of investigation (Ozdamar et al.1991,1992;Kalayaci and Ozdamar,1995; Ozdamar and Kalayaci,1998),(Webber et al.,1994,1996). Though the following facts:(i)numerous systems include simultaneous EEG-multichannel analysis(Ozdamar et al., 1991;Gabor et al.,1996;Park et al.,1998;James et al., 1999;Kobayashi et al.,1999);(ii)difficulties are often enhanced by the presence in the EEG of artifacts induced by eye blinking,movement activity or electrode displace-ment;and(iii)each group followed a personal strategy guided by their objective and their neural network tools approach;most of the ANN-based systems performed with a high level of accuracy(Ozdamar et al.,1991;Webber et al.,1996;Gabor,1998;Park et al.,1998;Tarassenko et al.,1998;Ko and Chung,2000).Positive results of the preliminary and prolonged studies should stimulate ANN research in epilepsy.11.Brain computer interfaceBrain computer interface(BCI),a new form of commu-nication using only EEG signals generated from different mental tasks without any other information source,has emerged.Its objective is to construct a system enabling severely physically disabled patients to communicate with their surroundings.Several groups focused their activity on this challenge using neural networks.Subjects are asked to performed mental tasks and the EEG is recorded during each task session and the neural network-based systems try to classify the EEG records into the correct mental task class.For10years,the G.Pfurtscheller’s group at the Department of Medical Informatics of Graz University of Technology in Austria has pursued BCI system research.One approach was to estimate the possibility of predicting the side of hand movements using EEG records prior to voluntary right or left hand movements(Pfurtscheller et al.,1992b;Masic and Pfurtscheller,1993;Peltoranta and Pfurtscheller,1994;Masic et al.,1996).In some studies (Pfurtscheller et al.,1992b;Masic and Pfurtscheller,1993; Flotzinger et al.,1994;Masic et al.,1996)classification rates were not very high(from51to83%).However classi-fication accuracies as high as85–90%were achieved in other studies(Peltoranta and Pfurtscheller,1994;Peters et al.,2001).In another approach,neural network-based systems were trained to classify movement intention of left and right indexfinger or the foot(Peters et al.1998) using EEG autoregressive model parameters.Correct recog-nition was achieved in83%of testing data.Other labora-tories have had promising results in building BCI systems utilizing neural networks(Anderson et al.,1995;Fukuda et al.,1995;Hiraiwa et al.,1997;Millan et al.,1998;Penny and Roberts,1999;Babiloni et al.,2000)with promising results.The future for neural network-based BCI systems is most promising.12.Other clinical applications12.1.Various clinical applications involving ANN-based systems are presentedANN-based systems were used to analysis continuous EEG to detect brain dysmaturity in newborns(Holthausen et al.,2000)and in children(Moreno et al.1995),while another study(on newborns)demonstrated the efficiency of neural networks in detecting hearing impairment through auditory evoked potential records(Sanchez et al.,1995). Using auditory event related potentials,some investigators were able to detect attention deficit hyperactivity disorder in children with a high classification rate(Dickhaus and Hein-rich,1997;Heinrich et al.,1999;Heinrich et al.,2001). Carried out on more than100patients,the analysis of visual evoked potential by ANN-based systems has been successful for discriminating between migraine,tension-type headache patients and normals(De Tommaso et al., 1997,1999),and for classification of neuroophtalmological disorders(Swiercz et al.,1997).Use of ANNs to classify head-injured(Gupta et al.1995),multiple sclerosis(Wu et al.1993,1994b),schizophrenic(Magdolen et al.,1996; Papadourakis et al.,1996;Sveinsson et al.,1997),or demen-ted(Anderer et al.,1994;Riquelme et al.,1996)patients from controls were also positive.Several groups have successfully built ANN-based systems for the discrimina-tion of other neurological disorders such as Huntington disease(Papadourakis et al.,1996;Jervis et al.,1999), Parkinson disease(Jervis et al.,1999),or Alzheimer disease (Polikar et al.,1997;Hibino et al.,2000;Petrosian et al., 2001)from normals.ANN Tools were also employed to discriminate depressed subjects from normals(MagdolenC.Robert et al./Clinical Neurophysiology113(2002)694–701697et al.,1996;Mitra et al.,1996)and to classify alcoholics (Klo¨ppel,1994).Only few attempts were performed in thefield of phar-macoelectroencephalography(Gevins et al.,1988;Echauz and Vachtsevanos,1994).These applications are representative of the potential and efficiency of neural networks for EEG data analysis in various clinical environments.13.ConclusionThis review demonstrates the importance that neural network studies have taken in medicine and biology invol-ving EEG signal processing.Positive results obtained in most applications presented show relevance for processing electroencephalograms.Of studies presented,some are no longer in use,others are pilot or preliminary studies with promising results,others are included in clinical projects in which neural networks are an essential component.Added to the development of neural network technology,these observations are encouraging for future‘EEG/neural networks’.AcknowledgementsThe authors wish to thank Robert M.Johnson(‘Bob l’Apache’)for his appreciable help with the English version of this article.ReferencesAbeyratne UR,Kinouchi Y,Oki H,Okada J,Shichijo F,Matsumoto K.Artificial neural networks for source localization in the human brain.Brain Topogr1991;4(1):3–21.Abeyratne UR,Zhang G,Saratchandran P.EEG source localization:a comparative study of classical and neural network methods.Int J Neural Syst2001;11(4):349–360.Algaver T,Smith TJ,Vijai F.The use of artificial neural networks in biomedical technologies:an introduction.Biomed Instrum Technol 1994;28:315–322.Allen R,Smith D.Neuro-fuzzy closed-loop control of depth of anaesthesia.Artif Intell Med2001;21:185–191.Anderer P,Saletu B,Kloppel B,Semlitsch HV,Werner H.Discrimination between demented patients and normals based on topographic EEG slow wave activity:comparison between z statistics,discriminant analysis and artificial neural network classifiers.Electroenceph clin Neurophysiol1994;91:108–117.Anderson CW,Devulapalli S,Stolz EA.EEG signal classification with different signal representations.In:Giros F,Makhoul J,Manalakos E, Wilson E,editors.Neural networks for signal processing V,Piscataway NJ:IEEE Service Center,1995.pp.475–483.Babiloni F,Carducci F,Cerutti S,Liberati D,Rossini PM,Urbano A, Babiloni parison between human and artificial neural network detection of Laplacian-derived electroencephalographic activity related to unilateral voluntary put Biomed Res2000;33:59–74.Bankman IN,Sigilliti VG,Wise RA,Smith PL.Feature-based detection of the K-complex wave in the human electroencephalogram using neural networks.IEEE Trans Biomed Eng1992;39(12):1305–1310.Bargiotti S,Bartolini F,Cappellini V,Mecocci A.EEG signal compression based on ADPCM and neural network predictors.Proc14th Colloque sur le Traitement du Signal et des Images,September13–16,Juan-les-pins,1993.pp.1287–1290.Battiti,Sartori A,Tecchiolli G,Tonella P,Zorat A.Neural compression:an integrated application to EEG signals.Proc Int Workshop on Applica-tions of Neural Networks to Telecommunications,Stockholm,1995.pp.210–217.Baumgart-Schmitt R,Hermann WM,Eilers R,Bes F.On the use of neural network techniques to analyze sleep EEG data.First communication: application of evolutionary and genetic algorithms to reduce the feature space and to develop classification rules.Neuropsychobiol 1997;36:194–210.Belenky G,Sing HC,Thomas ML,Shaham Y,Balwinski S,Redmond DP, Balkin TJ.Discrimination of rested from sleep-deprived EEG in awake normal humans by artificial neural network.Proc IEEE Int Conf Neural Networks,IEEE Piscataway,NJ,1994.pp.3521–3524.Berger H.Uber das elektrenkephalogramm des menschen.Arch Psychiatr NervKrankh1929;87:527–570.Bogacz R,Markowska-Kaczma U,Kozik A.Blinking artefact recognition in EEG signal using artificial neural network.Proc4th Conf Neural Network&Applications,Poland:Zakopane,1999.p.6.Bruha I,Madhavan bined syntax–neural net method for pattern recognition of evoked potentials.In:Janicki R,Koczkodaj WW, put and Information,North-Holland:Elsevier Science Publishers BV,1989.pp.361–369.Bruha I,Madhavan GP,Chong e of multilayer perceptron for recognition of evoked potentials.Int J Pattern Recog Artif Intell 1990;4(4):705–716.Carpenter GA,Grossberg S.A self-organizing neural network for super-vised learning,recognition,and prediction.IEEE Commun Mag1992 Sep;30:38–49.Clochon P,Perchey G,Couque C,Rebeyrolle H,Bloyet D,Etevenon P.Automatized classification by neural networks of EEG signals with artefact rejection.Proc14th Ann Int Conf IEEE Eng Med&Biol Soc,Lyon,France:Satellite Symosium on Neuroscience and Technol-ogy,1992.pp.51–55.De Tommaso M,Sciruicchio V,Bellotti R,Castellano M,Tota P,Guido M, Sasanelli G,Puca FM.Discrimination between migraine patients and normal subjects based on steady state visual evoked potentials:discri-minant analysis and artificial neural network classifier.Funct Neurol 1997;12(6):333–338.De Tommaso M,Sciruicchio V,Bellotti R,Guido M,Sasanelli G,Specchio LM,Puca F.Photic driving response in primary headache:diagnostic value testes by discriminant analysis and neural network classifiers.Ital J Neurol Sci1999;20:23–28.Dickhaus H,Heinrich H.EP parametrization and classification using wave-let networks–theoretical concept and medical application.Stud Health Technol Inf1997;43:541–545.Dumitras A,Murgan AT,Lazarescu V.A quantitative study of evoked potential estimation using feed forward neural network.In:Vlontzos J,Hwang J-N,Wilson E,editors.Proc of the1994IEEE Workshop on Neural Networks for Signal Processing Ermioni,Greece,Sept1994.pp.606–615.Durka P,Kiezyk R,Blinowska K.Neural networks and wavelet analysis in EEG artefact recognition.II Konferencja Siec Neuronowe CH Zasto-sowania szczyrk301V4V,1996.pp.6.Available at:http://brain.fuw.e-du.pi/~durka/pd_paper.html.Eberhart RC,Dobbins RW,Webber WRS.Casenet:a neural network tool for EEG waveform classification.Proc IEEE Symp Comput Based Med Sys,Minneapolis,MN,June25–27,1989.pp.60–68.Echauz J,Vachtsevanos G.Neural network detection of antiepileptic drugs from single EEG trace.Proc Electro94Int Conf Boston MA,May10–12, 1994.pp.346–351.Available at:/23783.html Elo P,Saarinen J,Varri A,Nieminen H,Kaski K.Classification of epileptic EEG by using self-organizing maps.In:Aleksander I,Taylor J,editors.Artificial neural networks2,Elsevier,1992.pp.1147–1150.C.Robert et al./Clinical Neurophysiology113(2002)694–701 698Emiliani GMM,Frietman EEE.Automatic classification of EEGs with neural networks.J Microelectron Syst Integr1994;1(1):41–62. Flexer A,Rappelsberger P,Dorffner G.Processing of EEG coherence maps with topologically correct feature maps.In:Eiselt M et al.,editors.‘Quantitative and topological EEG and EMG analysis’universitaets-verlag Jena,1995.pp.397–406.Flotzinger D,Pfurtscheller G,Neuper C,Berger J,Mohl W.Classification of non-averaged EEG data by learning vector quantization and the influence of signal preprocessing.Med Biol Eng Comput 1994;32:571–576.Freeman A,Skapura DM.Neural networks.Algorithms.Applications and programming techniques,Reading,MA:Addison-Wesley,1991. Fukuda O,Tsuji T,Kaneko M.Pattern classification of EEG signals using a log-linearized gaussian mixture neural network.Proc IEEE Int Conf Neural Networks,Perth Western Australia,27Nov–1Dec,1995.pp.2479–2484.Fung KSM,Chan FMY,Lam FK,Liu JG,Poon PWF.Visual evoked potential enhancement by artificial neural networkfilter.Biomed Mater Eng1996;6:1–13.Fung KSM,Chan FHY,Lam FK,Poon PWF.A tracing evoked potential estimator.Med Biol Eng Comput1999;37:218–227.Gabor AJ.Seizure detection using a self-organizing neural network:vali-dation and comparison with other detection strategies.Electroenceph clin Neurophysiol1998;107:27–32.Gabor AJ,Seyal M.Automated interictal EEG spike detection using arti-ficial neural networks.Electroenceph clin Neurophysiol1992;83:271–280.Gabor AJ,Leach RR,Dowla FU.Automated seizure detection using a self-organizing neural network.Electroenceph clin Neurophysiol1996;99: 257–266.Gevins AS,Smith ME.Detecting transient cognitive impairment with EEG pattern recognition methods.Aviat Space Environ Med1999;70(10): 1018–1024.Gevins AS,Stone RK,Ragsdale SD.Differentiating the effects of three benzodiazepines on non-REM sleep EEG spectra.Neuropsychobiology 1988;19:108–115.Gro¨zinger M,Ro¨schke J,Klo¨ppel B.Automatic recognition of rapid eye movement(REM)sleep by artificial neural networks.J Sleep Res 1995;4:86–91.Gulati,A,Jacobs D,Silan MN,Pope AT.Classification and prediction of electroencephalographic data with hybrid neural networks to analyze human-automation related-effects.Proc7th Int Conf Artif Intel& Expert Sys Appl November,1995.pp.449–454.Gupta L,Molfese D,Tammana R.An artificial neural network approach to ERP classification.Brain Cogn1995;27:311–330.Heinrich H,Dickhaus H,Rothenberger A,Heinrich V,Moll GH.Single-sweep analysis of event-related potentials by wavelet networks–meth-odological basis and clinical application.IEEE Trans Biomed Eng 1999;46(7):867–879.Heinrich H,Moll GH,Dickhaus H,Kolev V,Yordanova J,Rothenberger A.Time-on-task analysis using wavelet networks in an event-related potential study on attention-deficit hyperactivity disorders.Clin Neuro-physiol2001;112:1280–1287.Herrmann C.Symbolical reasoning about numerical data:a hybrid approach.Appl Intell1997;7:339–354.Hertz J,Kroch A,Palmer RG.Introduction to the theory of neural compu-tation,New York,Singapore,Milan,Paris:Addison&Wesley Publish-ing Company,1991.Hibino S,Hanai T,Nagata E,Matsubara M,Fukagawa K,Shirataki T, Honda H,Kobayashi T.An assessment system of dementia of Alzhei-mer type using artificial neural networks.In:Malmgren H,Borga M, Niklasson L,editors.Artificial neural networks in medicine and biol-ogy.Proc ANNIMAB-1Conference,Goteborg,Sweden13–16May 2000,Springer-Verlag2000.pp.180–185.Hiraiwa A,Uchida N,Shimohara K,Sonehara N.Neural network recogni-tion of electro-encephalogram patterns preceding voluntary movement.Electron Commun Jpn1997;80(7Part3):65–73.Hoffmann K,Leistritz L,Feucht M,Benninger F,Reiter D,Witte H.Iden-tification of the stimulated hemiretina in primary school children and adults based on left and right hemifield pattern reversal visual evoked potentials–a comparative study.Clin Neurophysiol2001;112:359–368.Holthausen K,Breidbach O,Scheidt B,Frenzel J.Brain dysmaturity index for automatic detection of high-risk infants.Pediatr Neurol2000;22: 187–191.Hopfield JJ.Neural systems and physical systems with emergent selective computational abilities.Proc Natl Acad Sci USA1982;79:2554–2558. Huang JW,Lu YY,Nayak A,Roy RJ.Depth of anesthesia estimation and control.IEEE Trans Biomed Eng1999;46(1):71–81.Huuponen E,Varri A,Himanen SL,Hasan J,Lehtokangas M,Saarinen J.Autoassociative MLP in sleep spindle detection.J Med Syst2000;24(3):183–193.James CJ,Jones RD,Bones PJ,Carroll GJ.Detection of epileptiform discharges in the EEG by a hybrid system comprising mimetic,self-organized artificial neural network,and fuzzy logic stages.Clin Neuro-physiol1999;110:2049–2063.Jando G,Siegel RM,Horvath Z,Buzsaki G.Pattern recognition of the electroencephalogram by artificial neural networks.Electroenceph clin Neurophysiol1993;86:100–109.Jervis BW,Garcia T,Giahnakis EP.Probabilistic simplified fuzzy ARTMAP(PSFAM).IEE Proc Sci Meas Technol1999;146(4):165–169.Jouny C.Electrophysiologie des e´tats de vigilance chez l’homme,France: Ph.D.Dissertation,University of Strasbourg,1997.p.201. Joutsiniemi SL,Kaski S,Larsen TA.Self-organizing map in recognition of topographic patterns of EEG spectra.IEEE Trans Biomed Eng 1995;42(11):1062–1068.Jung TP,Makeig S,Stensmo M,Sejnowski T.Estimating alterness from EEG power spectrum.IEEE Trans Biomed Eng1997;44(1):60–69. Kalayaci T,Ozdamar O.Wavelet preprocessing for automated neural network detection of EEG spikes.IEEE Eng Med Biol1995;Mar/ Apr:160–166.Klo¨ppel B.Classification by neural networks of evoked potentials.Neurop-sychobiol1994;29:47–52.Ko CW,Chung HW.Automatic spike detection via artificial neural network using raw EEG data:effects of data preparation and implica-tions in the limitations of online recognition.Clin Neurophysiol 2000;111:477–481.Kobayashi K,James CJ,Nakahori T,Akiyama T,Gotman J.Isolation of epileptiform discharges from unaveraged EEG by independent compo-nent analysis.Clin Neurophysiol1999;110:1755–1763. Kohlmorgen J,Muller KR,Rittweger J,Pawelzik K.Analysis of wake/ sleep EEG with competing experts.In:Gerstner W,Germond A,Hasler M,Nicoud JD,editors.Proc Int Conf Art Neural,LNCS1327,Berlin: Springer,1997.pp.1077–1082.Kohonen T.Self-organization and associative memory,Springer-Series in Information Sciences,8.Berlin:Springer-Verlag,1988.Koska M,Rosipal R,Konig A,Trejo LJ.Estimation of human signal detection performance from event-related potentials using feed-forward neural network model.Selected papers of2nd European IEEE Work-shop CMP’96,Boston:Birkhauser,1996.p.10.Krajca V,Principe JC,Petranek S.The dimensionality reduction and recon-struction of the multichannel EEG signals by self organized principal components analysis.In:Prochazka A,Uhlir J,Sovka P,editors.Signal analysis and prediction I,Eurasip,Prague:ICT Press,1997.pp.406–408.Krkic C,Roberts SJ,Rezek I,Jordan C.EEG-based assessment of anaes-thetic depth using neural networks.Proc IEE Colloquium Artif Intell Methods Biosignal Anal Apr1996;100:10/1–10/6.Lange D,Siegelmann HT,Pratt H,Inbar GF.Overcoming selective ensem-ble averaging:unsupervised identification of event-related brain poten-tials.IEEE Trans Biomed Eng2000;47(6):822–826.Laskaris N,Fotopoulos S,Papathanasopoulos P,Bezerianos A.Robust moving averages with Hopfield neural network implementation forC.Robert et al./Clinical Neurophysiology113(2002)694–701699。