ECL-tEEG: A HYBRID DEEP LEARNING MODEL FOR ESTIMATING LEARNERS’ EMOTIONS FROM EEG SIGNALS

Abstract

Recognizing emotions and assessing student engagement are important for improving instructional effectiveness and personalizing teaching. Previous studies have primarily used video or image data with deep learning models to analyze learners’ behavior and emotions. However, these approaches are sensitive to camera angles, lighting conditions, and ambiguous facial expressions, which may reduce the accuracy of emotion assessment. To address these limitations, this study proposes ECL-tEEG, a model that combines a convolutional neural network (CNN), an LSTM–GRU architecture, and a Transformer to analyze electroencephalogram (EEG) signals for emotion recognition and student engagement assessment. The model uses the CNN to extract spatial features, the LSTM–GRU architecture to capture temporal dependencies, and the Transformer to model long-range relationships through parallel processing. Its performance was evaluated against two comparison models: CNN-LSTM-GRU and CNN-Transformer. ECL-tEEG achieved an accuracy of 71%, outperforming both comparison models in classifying positive, neutral, and negative emotions. The findings demonstrate the potential of EEG-based emotion recognition and provide a basis for intelligent instructional support systems that can personalize learning and help teachers adjust their instruction using neural signals alongside behavioral observations.