Automatic interictal epileptiform discharge (IED) detection based on convolutional neural network (CNN)

Abstract

Clinical diagnosis of epilepsy significantly relies on identifying interictal epileptiform discharge (IED) in electroencephalogram (EEG). IED is generally interpreted manually, and the related process is very time-consuming. Meanwhile, the process is expert-biased, which can easily lead to missed diagnosis and misdiagnosis. In recent years, with the development of deep learning, related algorithms have been used in automatic EEG analysis, but there are still few attempts in IED detection. This study uses the currently most popular convolutional neural network (CNN) framework for EEG analysis for automatic IED detection. The research topic is transferred into a 4-labels classification problem. The algorithm is validated on the long-term EEG of 11 pediatric patients with epilepsy. The computational results confirm that the CNN-based model can obtain high classification accuracy, up to 87%. The study may provide a reference for the future application of deep learning in automatic IED detection.

Interrater agreement of annotations of epileptiform discharges and its impact on deep learning – A pilot study

“…Including studies using either or both of these measures, twenty-two relevant studies were found spanning the years 2016–2023. The median ACC is 0.90 (min: 0.71; max: 1.00) and the median AUC is 0.94 (0.76, 1.00) (Abou Jaoude et al, 2020;Antonaides et al, 2017; Cheng et al, 2022; Chung et al, 2023; Faghihpirayesh et al, 2021; Fürbass et al, 2020; Geng et al, 2021; Jeon et al, 2022; Jiang et al, 2023; Johansen et al, 2016; Kural et al, 2022; McDougall et al, 2023; Medvedev et al, 2019; Nejedly et al, 2023; Nhu et al, 2023; Thangavel et al, 2021; Thomas et al, 2020; Thomas et al, 2021; Tjepkema-Cloostermans et al, 2018; Wang et al, 2023; Wei et al, 2021; Zhang et al, 2023). In these studies, there is a large variation in data sizes, in the methods to calculate ACC (e.g., some used balanced ACC) and AUC (e.g., ROC vs. precision and recall), making comparisons difficult.…”