Methodical investigations for simultaneous detection and classification ofrolandic spike activity

In a previous study the authors showed that rolandic spikes are characteris ed by typical field distributions of the spectral parameters instantaneous power and instantaneous frequency [14]. According to localisation of the fo cus and lateralisation of instantaneous power seven topographic spike class es were determined visually and verified with a Neural Network classifier ( multi layer perceptron - MLP) [9]. Based on these results an algorithm for simultaneous detection and classification of rolandic spike activity was de veloped [7]. Aim of this study was to check the results of visual spike cla ssification by means of a global optimising cluster algorithm and to test a dditional classifiers - Linear Discriminant Analysis (LDA) and a Cascade Co rrelation net (CC) for topographic spike classification and their applicati on in the developed spike detection algorithm. Essentially, the results of cluster analysis confirmed the visual spike classification. The number of " correct" classifications of visually selected instantaneous power distribut ions of rolandic spikes (7 classes) and non-spike activities (alpha- and EM G-activities) of 10 Routine EEG records was nearly the same for the three c lassifiers LDA, MLP and CC. Routine EEG records of three further children c ontaining more than 900 spikes were used to compare the performance of the spike detection algorithm using LDA, MLP or CC with the results of visual s pike detection by two experienced electroencephalographers. The best result s were obtained with the MLP as classifier in the developed detection algor ithm. The number of "false/positive" detections was significant lower than when using LDA or CC.