Pattern recognition of epileptiform events in EEG signals using Wavelet Scalograms

This paper conducts a study on the use of Wavelet Scalograms as input of Artificial Neural Networks (ANN) for the automatic detection of epileptiform events in EEG signals. For this purpose, two Wavelet families, Daubechies and Coiflet, were tested totalizing 20 Wavelet functions. The ANNs used for simulations were Multilayer Perceptrons that were trained with using cross-validation technique and early-stopping to avoid under-fitting and over-fitting effects. The EEG database used for the simulations had 1,134 events between background activity, eye blinks, noise and epileptiform events (spikes and/or sharp waves). The events were divided in three groups, being 40% for ANN training, 30% for validation and 30% for ANN testing. After simulations the efficiency, sensitivity, specificity and AUC value of the networks were analyzed. The preliminary results showed that Daubechies 1 and Coiflet 1 presented the highest efficiency values (98%), Coiflet 5 and Coiflet 3 had the best sensitivity and specificity results, respectively, and the highest AUC value (0.99) was obtained with Coiflet 3.