Enhancing Depression Detection from Emotion EEG with Temporal-Spatial-Spectral Representation Learning

Major depressive disorder (MDD) significantly impairs psychosocial functioning and reduces quality of life. Developing reliable, objective biomarkers for depression diagnosis remains a critical yet challenging task. Since Electroencephalography (EEG) allows for the examination of brain activity patterns associated with MDD, this work leverages deep learning techniques on emotion EEG data to bridge the gap between observable depression symptoms and underlying neural signatures for objective depression detection. EEG recordings were collected from 33 depressed patients (DPs) and 40 healthy controls (HCs) in response to happy, neutral, and sad emotional stimuli. We propose a hybrid Emotion EEG CNN-Transformer model (EMOCT) for DP-HC classification. By combining CNN and Transformer blocks, EMOCT effectively captures temporal, spectral, and spatial features, enabling a more comprehensive representation of brain activity related to depressed or healthy mental states. Our extensive experiments demonstrate that EMOCT outperforms other models in DP-HC classification, achieving accuracies of 85.97%, 82.83%, and 85.25% for happy, neutral, and sad emotion EEG data, respectively. The results highlight EMOCT's potential as an effective and objective diagnostic tool for depression, paving the way for improved clinical assessment and management of the disorder.