Deep Learning Approaches for EEG-Based Emotion Recognition: A Systematic Review on DEAP and DREAMER

Emotions are a fundamental part of human life and an important factorinfluencing individuals' thoughts, behaviors, and decisions. In the context of human-computer interaction, emotion recognition is becoming increasingly important, particularly for enriching the user experience and enabling more naturalinteractions. As a key component of affective computing, this capability allowssystems to sense and adapt to users’ emotional states. Recently, there has beenan increase in emotion recognition studies conducted using EEG signals. However, the multidimensional and noise-sensitive nature of EEG signals makes manual feature extraction difficult, making deep learning (DL) based approaches advantageous. This study systematically analyzes DL approaches applied to thewidely used EEG-based emotion recognition datasets DEAP and DREAMER. Atotal of 185 scientific publications were evaluated in terms of fundamental stagessuch as signal preprocessing, feature extraction, and DL architectures. The findings reveal that the CNN architecture (N=129) is the most frequently used approach, followed by the LSTM (N=45) and GRU (N=16). Artifact removal, filtering, and normalization techniques are frequently applied in preprocessingsteps, while frequency-based methods (N=132) are predominantly preferred forfeature extraction. In the reviewed studies, various approaches have been employed depending on whether raw or preprocessed EEG signals were used andon model requirements. The study contributes to the literature by synthesizingcurrent DL practices in EEG-based emotion recognition and by identifying common methodological patterns that may inform future research in the field.