EEG spike detection with a Kohonen feature map.

Artificial neural networks are widely used for pattern recognition tasks. For spike detection in electroencephalography (EEG), feedforward networks trained by the backpropagation algorithm are preferred by most authors. Opposed to this, we examined the off-line spike detection abilities of a Kohonen feature map (KFM), which is different from feedforward networks in certain aspects. The EEG data for the training set were obtained from patients with intractable partial epilepsies of mesiotemporal (n=2) or extratemporal (n=2) origin. For each patient the training set for the KFM included the same patterns of background activity and artifacts as well as the typical individual spike patterns. Three different-sized networks were examined (15 x 15 cells, 25 x 25 cells, and 60 x 60 cells in the Kohonen layer). To investigate the quality of spike detection the results obtained with the KFM were compared with the findings of two board-certified electroencephalographers. Application of a threshold based on the partial invariance of spike recognition against translation of the EEG provided an average sensitivity and selectivity of 80.2% at crossover threshold (71%-86%) depending on the networksize and noise. Multichannel EEG processing in real time will be available soon. In conclusion, pattern-based automated spike detection with a KFM is a promising approach in clinical epileptology and seems to be at least as accurate as other more-established methods of spike detection.