Deep Learning for Epilepsy monitoring: A survey

E3S Web of Conferences 351, 01068 (2022) ICIES’22 https://doi.org/10.1051/e3sconf/202235101068 Deep Learning for Epilepsy monitoring: A survey Ghita Amrani,1, * Amina Adadi2, Mohammed Berrada1, and Zouhayr Souirti3,4 1 Laboratory of Artificial Intelligence, Data Science and Emerging Systems, National School of Applied Sciences of Fez, USMBA Fez, Morocco. 2 ISIC Research Team of High School of Technology, 2ISEI Laboratory, Moulay Ismail University, Meknes, Morocco. 3 Clinical Neurosciences Laboratory, Faculty of Medicine and Pharmacy of Fez, USMBA Fez, Morocco. 4 Neurology Department, Sleep center Hassan II University Hospital, USMBA Fez, Morocco. Abstract. Diagnosis of epilepsy can be expensive, time-consuming, and often inaccurate. The gold standard diagnostic monitoring is continuous video-electroencephalography (EEG), which ideally captures all epileptic events and dis-charges. Automated monitoring of seizures and epileptic activity from EEG would save time and resources, it is the focus of much EEG-based epilepsy research. The purpose of this paper is to provide a survey in order to understand, classify and benchmark the key parameters of deep learning-based approaches that were applied in the processing of EEG signals for epilepsy monitoring. This survey identifies the availability of data and the black-box nature of DL as the main challenges hindering the clinical acceptance of EEG analysis systems based on Deep Learning and suggests the use of Explainable Artificial Intelligence (XAI) and Transfer Learning to overcome these issues. It also underlines the need for more research to recognize the full potential of big data, Computing Edge, IoT to implement wearable devices that can assist epileptic patients and improve their quality of life. Keywords. Epilepsy, Deep learning, Electroencephalography, EEG signal analysis. 1 Introduction As defined by the International League Against Epilepsy, epilepsy is a temporary occurrence of signs and symptoms caused by abnormally synchronized and rapidly changing activity of neurons in the brain. [1]. About 50 million people worldwide are affected by this chronic neurological brain disorder [2] which is characterized by epileptic seizures that can lead to significant social, cognitive, physiological, and neurological consequences and may lead to death if not monitored and diagnosed appropriately [3]. The Electroencephalographic signals (EEG) can provide key information to identify neurological conditions and should be recorded to localize epileptic seizures. EEG is the most efficient method of recording and analyzing brain activity regardless of its sensitivity to noise, its key advantages are non-invasiveness, portability, costeffectiveness, relative ease of use, and exceptional submillisecond temporal resolution[4]. Diagnosing epilepsy using EEG signals is time-consuming and takes a lot of effort and it is prone to human error, as the neurologist has to carefully examine many hours of EEG signal recording. Computer-Aided Diagnosis (CAD) solutions are needed to assist neurologists and patients in the identification of seizures. Generally, a CAD system to EEG signal analysis follows a four steps process: (i) Signal acquisition: This step involves capturing electrical activity on the scalp using EEG recording methods. (ii) Preprocessing: since the raw EEG data are subject to artifacts and noise [5]. Such artifacts must be identified and eliminated using a set of manipulations for the subsequent processing steps. iii) Feature extraction: This step aims to analyze the preprocessed signal and decrease the number of features in it by creating new features from the existing ones which then should summarize most of the information in the original feature set. (iv) Feature classification: This step involves designing a properly structured and well-defined model for disease detection or prediction or pattern recognition in the signal. Automatizing the previous steps of EEG signal analysis using DL is an important stage towards building more * Corresponding author: © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 (http://creativecommons.org/licenses/by/4.0/). E3S Web of Conferences 351, 01068 (2022) ICIES’22 https://doi.org/10.1051/e3sconf/202235101068 practical EEG applications relying less on human professional capacities requiring longer training and expertise. The purpose of this paper is to provide a survey in order to understand, classify and benchmark the key parameters of deep learning-based approaches that were applied in the processing of EEG signals for epilepsy monitoring. In this sense, we make two main contributions:   Based on literature analysis, we give a structured overview of the existing approaches with respect of the four steps process of analyzing EEG signals. We spot and discuss future research avenues and trends in this field that we filtered from the survey of literature. The rest of this paper is organized as follows. A survey based on the epilepsy monitoring automation process is presented in section 2. The potential research directions and open problems are summarized in section 3. Finally, the conclusion is delineated in section 4. 2 Survey on DL for epilepsy: a processbased perspective In the quest to digitalize epilepsy monitoring, several DL methods have been proposed in the literature. Based on the conducted survey of the literature, we noted that the existing approaches vary mainly on the aim of using DL techniques, for example, to reduce noise, to augment data, to extract features, or to decode the neurologic disorder. Hence, we choose to arrange the included studies according to the stages of EEG analysis process described in the previous section. Prior to this paper, we know of few works that tried to study the issue from this standpoint. For example, the review of Shoeibi et al. [6] adopted a technical viewpoint and focused on the type of DL architectures, in addition, it focused only on epileptic seizure detection problems. 2.1. Data acquisition  Public datasets: Epilepsy databases are important in the design of accurate and robust CADs. Several EEG datasets were used by the included studies., namely, Freiburg [7], CHB-MIT [8], Intracranial EEG dataset Kaggle [9], Bonn [10], BernBarcelona [11], Epileptic seizure recognition dataset [12], TUH abnormal EEG corpus [13]. Although the number of data sets in this area is quite large, they cannot be mixed easily due to different frequency sampling and other parameters. In addition, many of the public datasets suffer from the quality of data.  Private datasets: Private datasets are gathered by scholars in their laboratory, hospital, or institution. While private datasets contain relatively a better quality of data, they remain limited in quantity. Additionally, private databases are not made available online so that other researc (...truncated)