Aprendizado profundo aplicado ao reconhecimento de emoções a partir de sinais de eletroencefalograma

Abstract

The present work is justified by the growing demand for non-invasive and efficient emotion recognition methods capable of balancing accuracy and computational cost. Emotion re-cognition using electroencephalogram (EEG) signals has gained prominence in applications such as brain-computer interfaces and mental health. However, many current deep learning models exhibit high computational complexity, which limits their practical use in edge computing devices or real-time systems. This study proposes and evaluates a simplified and efficient architecture, consisting of a pipeline structured in three main stages: data preprocessing and feature extraction through differential entropy calculation; spatial modeling of electrode relationships using a Graph Convolutional Neural Network (GCNN); and binary classification of valence and arousal dimensions using a Multi-Layer Perceptron (MLP) classifier. To mitigate the inherent imbalance of the dataset used, the method integrates class weights directly into the Cross-Entropy Loss function. The experiments were validated using the DEAP dataset under a rigorous subject-independent 5-fold cross-validation scheme. The results demonstrated mean accuracies of 66.21% for valence and 59.37% for arousal, accompanied by statistical stability evidenced by low standard deviations. The critical analysis indicates that class weighting allowed for a more realistic and balanced trade-off between the classifier’s sensitivity (recall) and specificity, overcoming the conservative bias of models without class treatment. It is concluded that the architecture fulfills the objective of providing a lean and methodologically consistent solution, serving as a reproducible and viable foundation for future implementations in real-time emotion recognition systems.