Automated seizure detection using a self-organizing neural network.

An algorithm for automated seizure detection using the self-organizing map (SOM) neural network (NN), with unsupervised training, was used to detect seizures in 24 long-term EEG recordings. The detection paradigm was tested on a constant 8 channel subset of 18 channel scalp EEG recordings. The NN was trained to recognize seizures using 98 training examples. A strategy was devised using wavelet transform to construct a filter that was 'matched' to the frequency features of examples used to train the NN. Four second epochs of training examples and EEGs being tested were transformed into time-independent representations of spectrograms resulting in a time-frequency representation of the time-series. Rule-based long and short term contextual features were used for detection in association with the NN. Fifty-six seizures were detected from a possible 62 (90%) associated with an average 0.71 +/- 0.79 false-positive errors per hour using the same 'population' detection parameters. When the sensitivity for detection was increased, all but one of the 62 seizures were detected (98%). Less than 1.0 false-positive error per hour occurred in all but 5 records when using the 'population' parameters. The combination of rule-based detection criteria employing contextual parameters and unsupervised training of NNs to recognize time-frequency patterns is a promising direction for automated seizure detection.