EmoDLNet: End-to-End Multi-Scale Spatio-Temporal Deep Learning for EEG-Based Emotion Recognition in Affective Human-Computer Interaction

Abstract Real-time emotion recognition from EEG signals is crucial for enhancing user engagement in human-computer interaction (HCI) applications, yet extracting emotion-specific features remains a challenge. This paper proposes EmoDLNet, a deep learning framework combining convolutional neural networks (CNNs) and hierarchical gated recurrent units (GRUs) to automate feature extraction from raw EEG data for emotion classification. We evaluate EmoDLNet on two datasets: DREAMER for binary arousal/valence classification and SEED-V for multi-class (happy, sad, fear, disgust, and neutral) emotion classification. On DREAMER, EmoDLNet achieved 89.4% (arousal) and 84.5% (valence) classification accuracy, surpassing state-of-the-art methods by 6.1% and 4.1%, respectively. On SEED-V, the model demonstrated robust multi-class performance, with an average accuracy of 89.9%, including 96.1% for disgust and 93.6% for happy emotions, outperforming baseline methods. Furthermore, the model’s inference latency of 67.10 ms per 1-second epoch shows its capability to support real-time emotion classification, making it suitable for brain-computer interaction (BCI) and HCI applications.