CIDDA: Classifier-driven Implicit Discriminator Domain Adaptation for EEG-Based Emotion Recognition and Depression Severity Grading

Variations in electroencephalogram (EEG) signals reflect the brain’s functional activity. In the context of the affective brain-computer interface, exploring whether emotion recognition knowledge can be used to assess depression severity may provide clues for further understanding of the relationship between depression and emotion. However, a pervasive challenge in EEG-based emotion recognition and depression severity grading is the variability in EEG recordings across subjects and sessions. Recent advancements have seen the application of deep domain adaptation methods aimed at learning domain-invariant features for cross-subject emotion recognition. However, these approaches have limitations, including ambiguous boundaries resulting from multi-classifier strategies, class information destruction, and mode collapse resulting from explicit domain discriminators and gradient reversal layers. So, a novel Classifier-driven Implicit Discriminator Domain Adaptation (CIDDA) method was proposed. This method has two key features: first, the employment of an implicit discriminator model design, which reduces both the number of parameters and computational costs; second, the integration of multiple optimization objectives through a joint classifier and the fusion of three types of loss functions, minimizing domain differences and maximizing label prediction accuracy. These features facilitate the alignment of marginal distributions and the reinforcement of conditional distributions about EEG features. Experiments on the public DEAP, SEED, and MODMA datasets, as well as our SignBrain depression EEG dataset validate the ability of our approach in emotion recognition and depression severity grading. These results demonstrate that our approach advances the potential for the clinical application of EEG signals.