214 Method Inform Med 1999; 38: 214 –24 Methods of Information in Medicine © F.K. Schattauer Verlagsgesellschaft mbH (1999) 1. Introduction Classification (i.e., detection and la- belling of episodes) and segmentation (i.e., discrimination between episodes) are common problems in biosignal pro- cessing. From a general point of view, both problems can be formulated as the finding of a mapping from one signal (the input) to another signal (either a sequence of labels for classification tasks, or a sequence of Boolean values for segmentation tasks), thus consti- tuting a specific filtering task. Neither the form nor the parameters of this mapping are usually known. Instead, labelled episodes of signals may be given that have been processed by a supervisor (e.g., a human expert or a sophisticated algorithm). If one can restrict the class of possible mappings to linear ones, determination of the optimal parameters is reduced to the construction of optimal linear filters, which is a standard task in signal pro- cessing that can be solved efficiently [1]. However, as the relation between input time series and target output is clearly non-linear, the linear filter approach is often inappropriate for solution of the problem classes described. As an alter- native, the linear filter with variable coefficients can be substituted by a more general class of nonlinear, para- metric systems. Among other methods, Neural Networks (NNs) have gained strong appeal during the last few years. In its simplest form, Time Delay Neural Network (TDNN), the NN consists of several input neurons that receive delayed samples from the input time series, of hidden neurons and of one output neuron (e.g. for segmentation tasks) or of several output neurons (e.g., for classification tasks), respec- tively. The labelled or segmented episodes then formulate a learning task for the NN. The only (but essential) difference from ordinary static learning tasks is that an ordering of the learning samples is given by time. This simple approach overcomes the limitations of linear filters. It does, however, suffer from several draw- backs: 1. It usually requires a large input di- mensionality: To characterize a segment, a number of samples have to be analyzed that depends on the spectral differences between segments of different clas- ses. Therefore, the solution of prac- tically relevant signal classification tasks often requires a relatively large input dimensionality (compared to those of static learning tasks). Be- sides an increased computational effort, this results in a large number of free parameters, and thus in gene- ralization problems [2, 3]. 2. It is not inherently invariant against translation: Let x 1 (n) be a periodical signal seg- ment x 1 (n) = x 1 (n + P 1 ) that has to be discriminated from a background noise or a mixture of other signals. In periods where x 1 (n) is present, the target output series o (n) is set to 1, and otherwise to 0. Then, the input dimension P 1 (length of the sliding A. Doering 1 , H. Jäger 1 , H. Witte 1 , M. Galicki 1 , C. Schelenz 2 , M. Specht 2 , K. Reinhart 2 , M. Eiselt 3 Adaptable Preprocessing Units and Neural Classification for the Segmentation of EEG Signals 1 Institute of Medical Statistics, Computer Science and Documentation, 2 Clinic of Anaesthesiology and Intensive Care, 3 Institute of Pathophysiology, Friedrich Schiller University Jena, Jena, Germany Abstract: In this contribution, a methodology for the simultaneous adaptation of preprocessing units (PPUs) for feature extraction and of neural classifiers that can be used for time series classification is presented. The approach is based upon an extension of the backpropagation algorithm for the correc- tion of the preprocessing parameters. In comparison with purely neural systems, the reduced input dimensionality improves the generalization capability and reduces the numerical effort. In comparison with PPUs with fixed parameters, the success of the adaptation is less sensitive to the choice of the parameters. The efficiency of the developed method is demonstrated via the use of quadratic filters with adaptable transmission bands as preprocessing units for the segmentation of two different types of discontinuous EEG: discontinuous neonatal EEG (burst-interburst seg- mentation) and EEG in deep stages of sedation (burst-suppression seg- mentation). Keywords: EEG Processing, Neural Networks, Classification, Intensive Care Downloaded by: Thieme E-Books & E-Journals. Copyrighted material.