Deep neural networks for the detection of temporal-lobe epileptiform discharges from scalp electroencephalograms

Abstract

Although interictal epileptiform discharges (IEDs) are a biomarker of epilepsy on electroencephalograms (EEGs), the manual annotation of IEDs is laborious. Thus, several IED detection methods have been proposed. However, the majority of these methods focus on single or whole cerebral channels. In this study, we examined the effect of channel selection and epoch length on the IED detection of temporal-lobe epilepsy (TLE). We identified two types of deep neural networks for IED detection: convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. We evaluated their performance using the F1-score, a global index of the sensitivity of IED detection, and the proportion of correctly detected IEDs based on scalp EEGs collected from 20 individuals with TLE. We then discovered that selecting EEGs from the affected temporal lobe or hemisphere was associated with a higher F1-score compared with selecting whole cerebral EEGs. In addition, we discovered that selecting a long EEG epoch (3 s) in the CNN model and selecting a short epoch (1.5 s) in the CNN + LSTM model resulted in the highest F1-score. In conclusion, deep neural networks are effective in detecting the TLE IEDs underlying an adequate epoch length when applied to EEGs of interest.