Search for a command to run...
Objective: Electroencephalography (EEG) is very crucial for understanding the dynamic healthy and pathological complex processes in the brain. However, the manual analysis of the EEG signal is very complex, time-consuming, and depends on the expertise and experience of the users. Hence, nowadays research is conducted on automated EEG signal analysis using artificial intelligence and computer-aided technologies. This would allow fast and highly accurate results. The goal of this paper is to provide an extensive review of the EEG signal analysis using deep learning (DL).Methods: This systematic literature review of EEG processing using Deep Learning (DL) was achieved on Web of Science, PubMed, and Science Direct databases, resulting in 403 identified papers. All collected studies were analyzed based on main disorders studied, type of tasks performed, data source, stages of analysis, and DL architecture.Results: DL in EEG processing is promising in various research applications. It covered the common neurological disorders diagnosis such as epilepsy, movement disorder, depression, schizophrenia, autism, alcohol use, attention, memory, sleep, pain, etc. The main tasks covered by the included studies are detection and classification. The average range of data sources utilized by the included studies is 127 subjects with an EEG recording a total duration of 458 hours. In fact, we identified the use of a plethora of DL architecture for EEG analysis. 57% of papers used Convolutional Neural Networks (CNNs), whereas Recurrent Neural Networks (RNNs) were the architecture choice of about 12% of papers. Combinations of CNNs and Long Short-Term Memory (LSTM) were used in 13% of studies. Generative Adversarial Networks (GAN) and Autoencoder (AEs) were used in 5% and 4% of papers respectively. Restricted Boltzmann Machine (RBMs), Deep Belief Networks (DBNs), and other hybrid architectures appeared in 1% of studies.Conclusion: This systematic review showed that DL is a powerful tool to process, analyze, and interpret EEG data without requiring extraction steps. These intelligent models can allow self-learning from the training data. On the other hand, DL models need a lot of data to learn, while suffering from a lack of confidence due to their black-box nature. Hence, studies on transfer learning and Explainable Artificial Intelligence (XAI) could help in solving such issues. Big Data, Cloud Computing, the Internet of Things (IoT), and closed-loop technology can also help DL-based systems in achieving fast, and accurate processing of EEG recordings.