Identification of inter-ictal spikes in the EEG using neural network analysis

Between seizures, the electroencephalogram (EEG) of subjects who suffer from epilepsy is usually characterised by occasional spikes or spike and wave complexes (inter-ictal activity). These are notoriously difficult to detect reliably, and they are occasionally missed by the clinicians who review the paper records retrospectively. The authors investigate the training and testing of a neural network classifier for the detection of inter-ictal spikes in these subjects. For the characterisation of the EEG signal, they consider both time-domain parameters normalised with respect to context and coefficients from an autoregressive model. It is shown how to use balanced databases to evaluate the discriminatory power of these parameters when they are used as the input features to a multi-layer perceptron (MLP). Both patient-specific classifiers and a generic system tested on independent test subjects are investigated. With the former, spikes are detected with an accuracy varying from 85.6% to 95.6%, a sensitivity varying from 83.1% to 97.3% and a specificity varying from 85.9% to 95.5%. The performance of the generic MLP system is not substantially degraded with respect to this, but there are too many false positives for the system to be considered for regular clinical use at the moment. The authors suggest how this problem might be solved using a combination of techniques.