UA-DAAN: An Uncertainty-aware Dynamic Adversarial Adaptation Network for EEG-based Depression Recognition

Depression is a common mental disorder characterized by symptoms such as a depressed mood, loss of interest, low self-esteem, and anxiety. Clinical diagnosis of depression often faces challenges due to the lack of objective indicators and the subjectivity of psychiatrists and patients. In recent years, with the rapid advancement of artificial intelligence technology, automatic depression diagnosis methods based on physiological signals have emerged. These methods have helped enhance the accuracy and objectivity of diagnosis. One such physiological signal used is the electroencephalogram (EEG), which is an easily obtainable, noninvasive, and cost-effective electrical signal recording the activity of neurons in the cerebral cortex. EEG is commonly used to observe brain states and diagnose mental illnesses. However, due to the high individual variability of EEG signals, existing methods often do not adequately address the issue of removing individual variability. Additionally, achieving high model reliability in disease recognition is crucial, but existing methods typically lack uncertainty estimation of recognition results. To tackle these challenges, this study introduces an uncertainty-aware dynamic adversarial adaptation network (UA-DAAN). This network incorporates adversarial learning concepts to address the significant individual variability in EEG data. It utilizes uncertainty-aware optimization of the dynamic domain adversarial process in a Bayesian neural network (BNN) to enhance the transferability of class-related features between source and target domains, improving the overall model's robustness, accuracy, and reliability. The experimental results strongly prove the effectiveness of this model in depression recognition.