CFAN: convolutional frequency-attention network for ECG-based emotion recognition

Emotion recognition using electrocardiogram (ECG) signals is a promising research direction with broad applications ranging from healthcare to human-computer interaction. However, mainstream convolutional neural network (CNN) and transformer-based models rarely exploit frequency-domain characteristics of ECG signals explicitly, limiting their cross-subject generalization capability. To address these limitations, we propose a convolutional frequency-attention network (CFAN) that integrates a frequency-aware attention mechanism with dynamic convolution to effectively extract and combine frequency and temporal features. CFAN comprises three key components: a frequency-aware attention module (FAAM), an attention-guided convolution neural network (AG-CNN) block, and a multi-layer perceptron (MLP) classifier, all working synergistically to enhance the efficacy of emotion recognition. We evaluate CFAN using leave-one-subject-out cross-validation by employing the WESAD dataset and further fine-tune the framework using data for individual subjects to reduce the inter-subject variability. CFAN outperforms state-of-the-art methods, achieving an accuracy of 76.06% and an F1-score of 0.75, providing an accurate and efficient solution for ECG-based emotion recognition.