[Discriminant analysis of graphic elements of the EEG. Application to the detection of episodes of diffuse spike-waves].
A database of EEG information was collected from EEG recordings performed in epileptic patients with diffuse spike-wave complex discharges. Normal activity, spike-waves, slow waves and artefacts were mixed up in these recordings. The analysis of EEGs stored in the database was performed, channel by channel, through a 2.56 s moving window. For each so defined EEG sequence, a set of 22 variables chosen for their discriminatory power was computed. A subset of 8 highly discriminating variables was selected by the means of a stepwise discriminant analysis. Each class of the learning set contained 40 up to 100 EEG sequences. A classifying algorithm that takes into account zones of uncertainty is proposed. It has been evaluated on a test set which was composed of 1981 EEG sequences issued from 15 different patients. The results have been checked by two neurologists. The agreement rate between each of them and the proposed algorithm was more than 92%; this result is comparable to the agreement rate between the two neurologists (94%). A contextual analysis algorithm, using bi-dimensional smoothing techniques, allowed to improve the agreement rates which exceeded 94%.