ReBAN-Net: An interpretable hybrid deep learning framework for ECG-based emotion recognition

Emotional assessment is essential for mental health management, yet questionnaire-based scales are subjective and facial- or voice-based approaches can be influenced by intentional behavior and environmental interference. This study proposes ReBAN-Net, a hybrid architecture integrating a one-dimensional residual convolutional network, a bidirectional long short-term memory network, and a soft attention mechanism for ECG-based emotion recognition. Three classification tasks were defined throughout the study: nine-class valence classification (levels 1-9), nine-class arousal classification (levels 1-9), and four-class affective-state classification (baseline, stress, amusement, and meditation). The model was evaluated on the WESAD dataset using a strict 15-fold leave-one-subject-out protocol. It achieved accuracies of 86.03% for nine-class valence, 86.27% for nine-class arousal, and 87.58% for four-class affective-state classification, with corresponding macro-AUC values of 96.07%, 96.24%, and 95.09%. Temporal class activation and attention-feature analyses show that the model focuses on physiologically meaningful ECG regions. These results demonstrate the feasibility and interpretability of ReBAN-Net and indicate its potential for subject-independent ECG-based emotion recognition.