Major depressive disorder detection via temporal-frequency-spatial transformer with sub-domain knowledge alignment using EEG.

Major Depressive Disorder (MDD) is a common mental illness that seriously jeopardizes the physical and mental health of patients. Accurate detection of MDD is crucial for treatment. Currently, there are significant differences in the EEG signals of each MDD patient, leading to lower accuracy of cross-subject MDD detection. Transformer-based methods have been used by scholars to detect MDD using electroencephalogram (EEG) data, but these methods often neglect the frequency features, focusing primarily on global domain adaptation (DA) while ignoring sub-domain alignment, resulting in loss of fine-grained discriminative information. To address this, we incorporate fine-grained frequency features to improve sub-domain alignment in DA rather than relying solely on global feature alignment. Building on the above analysis, we propose the TFST-SDKA model, a temporal-frequency-spatial Transformer (TFST) integrated with a sub-domain knowledge alignment (SDKA) method for MDD detection. The SDKA module classifies subjects into distinct sub-domains based on their labels by extracting fine-grained discriminative information from each subject. This process helps bridge the gap between source and target domains, enhancing the model's generalization. In addition, we propose a frequency attention (FA) mechanism, which uses discrete cosine transform (DCT) to convert EEG feature maps into the frequency domain. The FA extracts multiple frequency information of EEG signals associated with MDD and combines these frequency data to enhance the model's representational capability. As a result, the TFST-SDKA model improves EEG feature representation and aligns source and target domain features. Extensive experiments conducted on the MODMA and PRED+CT datasets demonstrate that our proposed TFST-SDKA model outperforms state-of-the-art (SOTA) methods in MDD detection tasks. Specifically, our method exceeds the SOTA methods by 1.42 % on the MODMA dataset and 1.16 % on the PRED+CT dataset in terms of accuracy.