The design and evaluation of an artificial neural network system for the detection of epileptogenic spikes is described. The system is composed of smaller neural network modules which are trained individually and organized in two levels. The first-level modules are trained to recognize candidate spikes in single referential electroencephalogram (EEG) channels. Original digitized data with a running window of 100 ms provided the input for the first-level modules. A second-level module is used for the spatial integration of 16 first-level modules. The system was trained and tested using clinical EEG data interpreted by four expert electroencephalographers. The results show that spikes can be recognized directly from unprocessed EEG and a second-level neural network can integrate spatial EEG information and eliminate false detections.