In this study, adaptive resonance theory (ART2) neural networks are investigated for on-line unsupervised recognition of EEG.spikes for epilepsy monitoring. ART2 networks are unsupervised self-organizing systems that cluster data into different classes. Learning is completed when these classes are labelled bv an expert and are saved in a look-up table for use by the system. Recognition of new data is accomplished by finding the class assigned by ART2 and searching through the table to get the proper labels. Unrecognized inputs are put into a new class for future labelling. In this study an ART2 neural network with 20 inputs was developed and trained using EEG data containing spike and non-spike waveforms. For comparison a 20 input multilayer perceptron was constructed and evaluated similarly. Evaluation with three sets of EEG patterns indicates that the ART2 network's performance is close to that of multilayer perceptron showing its high potential. Considering that ART2 can be trained with one or a few iterations (compared to thousands required for hackpropagation networks), monitoring systems with on-line training can be easily constructed. Such systems are highly desirable for long-term EEG monitoring due to large variations of spike waveforms that occur among individual patients and during long recording time periods.