Scalp EEG epileptogenic zone recognition and localization based on long-term recurrent convolutional network

Abstract

The scalp electroencephalogram (EEG), a non-invasive measure of brain's electrical activity, is commonly used ancillary test to aide in the diagnosis of epilepsy. Usually, neurologists employ direct visual inspection to identify epileptiform abnormalities. Therefore, electroencephalograms have been an essential integral to the researches which aim to automatically detect epilepsy. However, it is difficult because seizure manifestations on scalp EEG are extremely variable between patients, even the same patient. In addition, scalp EEG is usually composed of large number of noise signals which might cover the real features of seizure. To this challenge, we construct an 18-layer Long-Term recurrent convolutional network (LRCN) to automatic epileptogenic zone recognition and localization on scalp EEG. As far as we know, we are the first to train a deep learning classifier to identify seizures through the EEG images, just like neurologists direct visual inspection to identify epileptiform abnormalities. Furthermore, unlike the traditionally methods extracted features from channels manually, which neglected the association of brain's epileptiform abnormalities electrical transmission, seizures is considered as a continuous brain's abnormal electrical activity in our algorithm, from produce at one or several channels, transmission between channels, to flat again after seizures. The method was evaluated in 23 patients with a total of 198 seizures. The classifier shows reasonably good results, with 84% for sensitivity, 99% for specificity, and 99% for accuracy. False Positive Rate per hours exceeds significantly previous results obtained on cross-patient classifiers, with 0.2/h.