Interictal Epileptiform Discharge Detection Using Multi-Head Deep Convolutional Neural Network

Interictal epileptiform discharges (IEDs) are intermittent electrophysiological events that occur in patients with epilepsy between seizures. Automated detection of IEDs helps clinician to identify cortical irritations and relations to seizure recurrence. It also reduces the necessity of visual inspection by physicians interpreting the EEG. This paper presents a novel deep learning-based approach that combines one-dimensional local binary pattern symbolization method with a regularized multi-head one-dimensional convolutional neural network to learn unique morphological patterns from different EEG sub-bands for IED detection. Experimentation using the Temple University Events corpus scalp EEG data shows promising performance, e.g. F1-score of 87.18%.

I. INTRODUCTION

Interictal epileptiform discharges (IEDs), also known as epileptiform spikes, are electrophysiological events that occur between seizures in patients with epilepsy. IEDs indicate cortical irritability, may represent the seizure onset zone, and are seen with increased likelihood of seizures following the first seizure [1]. These oscillations are different from ictal discharges which represent seizure activity and are considered as non-ictal. However, as they can provide important insight in the diagnosis of epilepsy, capturing these transients can be valuable for epilepsy diagnosis. Visual inspection of IEDs from electroencephalogram (EEG) is arduous and inefficient, requires skilled neurophysiologists, and is prone to inter-observer error, which leads to a high misdiagnosis rate. Hence, developing robust and automated algorithms for IED detection and quantification from scalp or intracranial EEG (iEEG) can: (i) help understand the extent of cortical irritability in seizure patients, and may indicate possible seizure onset zone, (ii) support epilepsy diagnosis by predicting seizure recurrence, (iii) assist in closed-loop neurostimulation as well as online seizure monitoring, and (iv) reduce misdiagnosis rate, as well as the resources and time spent on visual analysis.