Detection of epileptogenic sharp transients using supervised and unsupervised artificial neural networks

The goal of this study was to develop, test, and compare two automated epileptogenic sharp transient (ST) detection systems based on two different types of artificial neural network (NN) architecture; a supervised backpropagation (BP) and an unsupervised adaptive resonance theory (ART2) network. The widely used BP NN was first investigated to use the results as performance reference. The BP NN architecture was modified in order to detect negative STs for a fair comparison with the ART2 architecture. An unsupervised ART2 NN, a previously unexplored architecture for this application, was the second system developed.; In contrast to most previous studies which used EEG extracted parameters, this study used raw EEG as input to both NN systems. Both architectures were trained and tested using large sets of exemplars (1477 and 1544 for two training/testing sets, TR1 and TR2, respectively) generated using one hour recording from 13 patients. The accuracies of the systems were compared with that of three EEGers which labeled all the recordings.; In this study, NNs which are non-algorithmic and non-rule-base approaches were successfully used for single channel ST detection. Both systems were first tested using different EEG events from the same 13 patients. The sensitivity (98%) and selectivity (97.5%) obtained for BP NNs were slightly better than the ones obtained for the ART2 NNs (sensitivity = 95.6%; selectivity = 97%) when small input window sizes were used. This difference increased as the number of input points increased beyond 25 points. These two architectures were later tested using the sliding window technique with 3 patients not included in the original training sets. This testing was done to simulate the real-time operation and capabilities of these networks. The FN ratio obtained for both networks were compared with that of three EEGers. However, the FP ratio remains a problem to be resolved.; The adaptive ART2 NN architecture offers a possibility of building an ST detection system that can perform on-line learning. Such a system is highly desirable for long-term EEG monitoring due to large variations of ST waveforms that occur among individual patients.