> *,)YbjbjWW ==]"""""""66666B6ZZZZZZZZ
$
"ZZZZZ
T ""ZZZT T T Z"Z"Z
66""""Z
T vT
""
ZN@-j;66j
Feature Extraction in Epilepsy by Cellular Nonlinear Networks (CNN)
Final Report
by PD Dr. Ronald Tetzlaff
My expectations concerning the working conditions in the Analogic and Neural Computing Laboratory, Computer and Automation Research Institute, of the Hungarian Academy of Sciences in Budapest, have been significantly excelled. I found perfect working conditions and help in an excellently organized institute, which is a real Center for Excellence. Besides several fruitful and interesting discussions we have started research collaboration, published a paper, and sent a pre-proposal (see enclosure) to the Future and Emerging Technologies programme of the European Commission. Especially, the below research results described in short show that a significant improvement has been achieved for the realization of an epilepsy warning and preventing system.
The analysis of invasive recorded brain electrical activity (BEA) for the realization of an epilepsy warning and preventing system for epileptic seizures has been treated between May and September 2001. As outlined in the enclosed working plan, the aim of these investigations is to analyze and verify recently introduced CNN algorithms for the extraction of BEA features. It is important to note that these features should reflect undergoing changes in the brains state of a patient, which could be regarded as precursors of an impending seizure occurring several minutes before the seizure onset. The considered CNN algorithms have been analyzed for several recordings on 4 patients. The obtained results are discussed in short in the following space. Intermediate results were presented at BIONICS Workshop (June 19-21, 2001) in Brussels and were published in the proceedings of the workshop. A manuscript giving a detailed discussion of all results is under preparation and will be submitted to Physical Review Journal for a publication as a regular paper.
Prediction of BEA by Using CNN
A discrete-time CNN (DTCNN) with non linear delay-type templates has been considered for the prediction of BEA. Thereby, for each of the successive data segments autonomous networks with the nearest neighborhood interactions have been determined in a supervised optimization procedure. The DTCNN have the structure of the used electrodes and the elements of the feedback templates, which are nonlinear functions of the cell outputs, have been respresented by third order polynoms. The detection of distinct changes of a DTCNN predictor is the aim of these investigations. Therefore, the parameter of the DTCNN, which are the polynomial coefficients of the feedback functions, have been analyzed for successive signal segments. The first results obtained from the data of one patient have been presented in the above mentioned BIONICS workshop. Meanwhile, the analysis has been performed for the complete databasis leading to the following results: Distinct fluctuations of the coefficients of the second and third order terms can be observed only some minutes before the onset of an epileptic seizure for two patients. These fluctuations, which do not occur in any other recording without a seizure, could be regarded as precursors of an impending seizure and simply detected using a threshold function.
Considering the results obtained for the recordings of the other 2 patients, such fluctuations occur at the seizure onset for one patient, whereas hereby the recording of BEA has been started shortly before the onset. Therefore, it is an open question in this case whether such fluctuations occur some minutes before the seizure onset. Regarding the results obtained for another patient, it seems to be that these fluctuations are not relevant. It is planned to verify these new promising exciting results for recordings of other patients to be later used under practical circumstances.
The Pattern Detection Algorithm
The aim of these investigation is the determination of an algorithm for the detection of certain binary BEA patterns occuring in successive signal segments. Thereby, distinct changes of the frequency of occurences of such patterns have been of particular interest. Firstly, binary BEA signal segments have been obtained for different threshold level values l = {-1,....,1}, whereas one signal segment denotes the initial condition of a standard CNN having linear weight functions. Then, as outlined in [1], the frequency of occurence of all possible 512 patterns with 3 x 3 binary values have been determined for successive signal segments considering a certain electrode of a multichannel recording. While first results have been presented in [1] meanwhile the analysis has been performed for the complete data basis leading the following results. In all cases, distinct changes of the frequency of occurence which can be observed only some minutes before the onset of an epileptic have been detected. Certain patterns occur frequently not exceeding a maximum time distance between the occurence of two successive patterns. Only some minutes before the onset of an epileptic seizure, these binary pattern do not occur again, exceeding the maximum time distance which has been observed until then. A CNN algorithm has been developed and tested for the detection of such distinct changes for the anticipation of epileptic seizures.
[1] Tetzlaff, R., Wei, D., and Kunz, R.:Anticipation of Epileptic Seizures by Cellular Neural Networks (CNN) Algorithms?; Proc. of the Bionics workshop, Brussels 2001, in print
ljCJmHmHCJ5CJDQklij=>
&F8P$dh$$dhDQklij=>. A!"#$%
[$@$NormlmH@A@@Bekezds alap-betqtpusaLC`LSzvegtrzs behzssalCJmH nH<B`<SzvegtrzsdhCJmHnH
!"*+45=DIJPQS[abjlQRinvz
!$%.578@ACDGHPQXY]_bcfjop~
$%,-/034;<AGNPUV[_ghjkuy{|
%&-.12;<>?DEGHKLUV[]ijqrvw
$(-.9:=>BFOPRS[\bknq|}QRQRQS=>QTlh<>
MTA SZTAKIED:\CoE-kislemezekrol copyzva\Finalrep-RonaldTetzlaff-May-Sept2001.doc
MTA SZTAKI[\\IBOLYA\pendrive (E)\COE-reports\Visitors-reports\Finalrep-RonaldTetzlaff-May-Sept2001.doc@p@[Times New RomanThorndale5Symbol3&Arial"pׄׄ %20CFeature Extraction in Epilepsy by Cellular Nonlinear Networks (CNN)
MTA SZTAKI
MTA SZTAKI
Oh+'0 ,8
T`
lxDFeature Extraction in Epilepsy by Cellular Nonlinear Networks (CNN)eatMTA SZTAKIrTA NormalAMTA SZTAKIr2A Microsoft Word 8.0i@@2oj;@2oj;
՜.+,D՜.+,x4hp
MTA SZTAKI% jDFeature Extraction in Epilepsy by Cellular Nonlinear Networks (CNN)Title 6>
_PID_GUIDAN{07DC94C2-A76D-11D8-A7E6-00C04FB40E59}
"#$%&'(+Root Entry Fj;pj;-1TableWordDocument SummaryInformation(DocumentSummaryInformation8!CompObjlObjectPoolpj;pj;
FMicrosoft Word dokumentum
MSWordDocWord.Document.89q