Depression is primarily characterized by emotional dysfunction, reflected in heightened negative emotions and diminished positive emotions. Consequently, accurate emotion recognition has become a crucial approach for assessing depression. Among the various signals employed for emotion recognition, electroencephalogram (EEG) signals have gained considerable attention due to their rich spatiotemporal information across multiple channels. In this study, EEG data were first preprocessed using filtering and Euclidean alignment. For feature extraction, time-frequency features were obtained through short-time Fourier transform and Hilbert–Huang transform, while spatial features were extracted using convolutional neural networks. Temporal relationships were then explored using a bi-directional long short-term memory network. Prior to convolution operations, EEG features were transformed into 3D tensors based on the unique topology of EEG channels. The proposed framework demonstrated strong performance on two emotion datasets: SEED and the Emotional BCI dataset from the 2020 WORLD ROBOT COMPETITION. When applied to depression recognition, the method achieved an accuracy exceeding 70% under five-fold cross-validation. Moreover, using a subject-independent protocol on the SEED dataset, the approach attained state-of-the-art performance, surpassing previous studies. Overall, we present a novel EEG-based emotion recognition framework for depression detection, offering a robust algorithm suitable for real-time clinical applications.
