EEG-Based Emotion Prediction for Non-Verbal Patients Using Dynamic Graph Convolutional Neural Networks

Abstract

Emotion Recognition using electroencephalography (EEG) signals provides a new avenue in developing non-invasive emotion characterization techniques for patients who lack verbal communication skills, including cases of impaired consciousness, severe autism, locked-in syndrome, or post-stroke aphasia in affective computing and brain computer interface (BCI) systems. In this regard, a Dynamic Graph Convolutional Neural Networks (DGCNN)-based framework for recognizing emotions is introduced. In particular, within the presented DGCNN-based framework, functional connections between EEG channels are inferred in a data-driven manner. For the feature extraction, Differential Entropy with Linear Dynamical System (DE_LDS) approach is adopted, which analyzes five frequency bands $(\delta, \theta, \alpha$, $\beta, \gamma)$ of 62 channels of EEG signals, wherein the signal data points represented as a graph of electrode sites. A learnable adjacency matrix for the proposed framework has been further regularized by incorporating the graph Laplacian smoothness constraints. To strictly prevent data leakage, a trial-level train-validationtest split was implemented on the SEED dataset. As a result, the DGCNN achieved an accuracy of 47.7%, outperforming the SVM baseline (43.1%) and an EEG Transformer (41.9%). Furthermore, a Domain Adversarial Neural Network (DANN) extension mitigated cross-session variance, raising accuracy to $\mathbf{4 8. 9 \%}$, while CPU benchmarking confirmed a real-time latency of 0.0744 ms per window.