EEG-Based Emotion Recognition: A Systematic Review of Traditional and Deep Learning Methods

OBJECTIVE This systematic review provides a synthesis of the existing data concerning the Electroencephalogram (EEG)-based emotion recognition and assesses the development of the old machine learning models to the current deep learning models. The purpose of the review is the comparison of their performance and the identification of trends in the approaches to the methodology and the evaluation of the strength and the reproducibility of the discipline. METHODS The review was done based on the Preferred Reporting Items of Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines. Five electronic databases (IEEE Xplore, Scopus, PubMed, ScienceDirect, and SpringerLink) that have been published not earlier than January 2012 were searched systematically. Due to the removal of duplicates and two rounds of screening against pre-defined inclusion criteria, 50 studies were incorporated to be final synthesized. FINDINGS It has been demonstrated that there is a definite trend towards end-to-end deep learning models, especially Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and architectures combining both. These models, particularly when using maps of topographic features and maps of functional connectivity, have shown a higher level of performance, and their classification accuracies of 90 percent or higher in benchmark data such as DEAP and SEED in subject-dependent contexts. But, there is a significant decline in the performance in cross-subject validation, which is an outstanding generalization issue. It also becomes evident during the synthesis that validation protocols, data preprocessing and reporting standards exhibit high heterogeneity, thus making it difficult to directly compare them and jeopardizing reproducibility. CONCLUSION Deep learning approaches are an important development in emotion recognition of EEG, but the area is plagued by lack of uniformity and focus on real-world applicability. The next step in work is to focus on the creation of standardized evaluation metrics, explicable AI methods, and effective, cross-subject models to enable the movement of laboratory studies to the reliable, deployable systems.