MS-DAAN: A Multi-Source Dynamic Adversarial Adaptation Network for EEG-Based Depression Recognition

Depression has become one of the most prevalent mental health disorders worldwide, highlighting the urgent need for objective and reliable auxiliary diagnostic methods. Electroencephalography (EEG), as a non-invasive technique with high temporal resolution, shows great promise in depression recognition. However, the significant inter-subject variability inherent in EEG signals limits the generalization ability of traditional machine learning and deep learning models in cross-subject scenarios. Although incorporating multi-source data can enhance the representational capacity of transfer learning, it also introduces new challenges, such as distributional conflicts and adaptation strategy inconsistencies between sources, which can lead to negative transfer. To address these issues, we propose a Multi-Source Dynamic Adversarial Adaptation Network (MS-DAAN). This framework constructs independent feature extraction and adversarial adaptation branches for each source domain, incorporates an unsupervised EEG-based source clustering mechanism to form semantically coherent subdomains, and introduces a target-guided source attention module to dynamically weight each source according to its statistical similarity to the target domain. Experimental results demonstrate that MS-DAAN significantly outperforms existing methods across multiple evaluation metrics, validating its effectiveness and robustness in cross-subject EEGbased depression recognition.