CSDA: Cross-Session Domain Adaptation in Auditory Attention Decoding of EEG for a Single Subject

Auditory attention decoding endeavors to extract specific features from human brain signals, such as electroencephalogram, to identify the target speaker. Harnessing the human brain's extraordinary capability to perform cocktail party effects, we are introducing a pioneering single-subject cross-session (SCS) task. This task holds immense potential in enhancing the adapt-ability of hearing assistance devices to diverse real-world environments. Acknowledging the variability and instability of EEG signals across different scenarios, which often leads to a significant decline in model performance when validated in diverse domains, we propose a novel cross-session domain adaptation (CSDA) approach. Specifically, CSDA utilizes domain-specific feature representations and classifier predictions, coupled with a domain discriminator, to minimize cross-domain discrepancies and enhance model performance. The experimental results demonstrate significant improvements in task performance. Moreover, our extensive generalization experiments indicate that this domain adaptation method outperforms other pre-existing approaches in the SCS task, further validating the effectiveness of CSDA. The introduction of this novel task-type not only showcases the innovation and practicality of our research but also holds the potential to drive significant advancements in human-computer interaction for neuro-steered hearing assistance devices, ultimately enhancing their applicability in real-world scenarios.