O20 Deep learning with convolutional neural networks for detection of interictal epileptiform discharges

Abstract

Objectives Electroencephalography is a standard clinical procedure for diagnosing epilepsy, and visual EEG analysis is gold standard. We propose a novel computer-assisted approach for interictal epileptiform discharge (IED) detection using deep learning with convolutional neural networks. Methods IEDs of 19-channel routine EEG recordings from 57 patients diagnosed with focal epilepsy were annotated by experts. Forty-seven of these and 46 normal EEGs were used as the training set. An independent test set was created from the remaining ten abnormal and ten normal EEGs. Each annotated IED was extracted as a 2-s EEG epoch in a bipolar montage, while 2-s non-IED epochs came from normal EEGs. All epochs were processed after band-pass filtering at 0.5–30 Hz. The deep learning architecture was implemented in Python using Keras and GPU and comprised of two convolutional and three fully-connected layers. Results The training set consisted of 2030 epochs with IEDs and 2995 normal epochs; the test set contained 491 and 5740 epochs, respectively. The best network had 64 different filters in the first convolutional layer and 128 in the second. Evaluation in the test set resulted in a sensitivity of 0.92 and specificity of 0.97. Discussion In this pilot, convolutional neural networks showed ability to detect IEDs in raw EEG epochs with high sensitivity and specificity. Further tuning of model parameters may improve performance which will be evaluated on a larger dataset. Conclusions Deep learning allows detection of IEDs in raw EEG epochs. Significance Deep learning has promise to assist visual assessment in detecting IEDs.