Graph attention network para reconhecimento de emoções a partir de eletroencefalogramas

Abstract

The automatic recognition of emotional states is relevant in several areas, such as mental health, brain-computer interfaces, and affective monitoring systems. Electroencephalography (EEG) is a non-invasive technique for recording brain electrical activity that stands out in this context for directly reflecting the neural states associated with emotions, being harder to mask than facial expressions or voice. However, characteristics such as the non-stationary nature of the signals, low signal-to-noise ratio, and inter-subject variability make this task challenging. In this scenario, Graph Neural Networks (GNNs) stand out for directly modeling the relationships between EEG channels. Thus, this work proposes a model based on Graph Attention Network (GAT), a GNN variant that adaptively weights the contribution of each neighboring channel during information aggregation, for the classification of three emotional states (positive, neutral, and negative) present in the SJTU Emotion EEG Dataset (SEED). Each EEG sample is represented as a graph whose nodes correspond to EEG channels and whose edges connect pairs of channels whose cosine similarity between their respective feature vectors exceeds the 70th percentile of the sample similarity distribution. Among the 62 channels and 5 frequency bands conventionally used in the literature, only 4 channels (FT7, FT8, T7, and T8) and 2 bands (delta and theta) were selected, resulting in a compact configuration with potential application in wearable devices with a limited number of electrodes. Experiments were conducted under the Leave-One-Subject-Out (LOSO) protocol, which evaluates generalization across subjects. The model achieves a mean accuracy of 95.38% over 10 independent runs, outperforming many comparable works that use the full channel and band configuration, indicating that graph-based modeling captures relevant discriminative patterns even with a reduced configuration.