TFDISNet: Temporal-frequency domain-invariant and domain-specific feature learning network for enhanced auditory attention decoding from EEG signals

Auditory Attention Decoding (AAD) from Electroencephalogram (EEG) signals presents a significant challenge in brain-computer interface (BCI) research due to the intricate nature of neural patterns. Existing approaches often fail to effectively integrate temporal and frequency domain information, resulting in constrained classification accuracy and robustness. To address these shortcomings, a novel framework, termed the Temporal-Frequency Domain-Invariant and Domain-Specific Feature Learning Network (TFDISNet), is proposed to enhance AAD performance. A dual-branch architecture is utilized to independently extract features from the temporal and frequency domains, which are subsequently fused through an advanced integration strategy. Within the fusion module, shared features, common across both domains, are aligned by minimizing a similarity loss, while domain-specific features, essential for the task, are preserved through the application of a dissimilarity loss. Additionally, a reconstruction loss is employed to ensure that the fused features accurately represent the original signal. These fused features are then subjected to classification, effectively capturing both shared and unique characteristics to improve the robustness and accuracy of AAD. Experimental results show TFDISNet outperforms state-of-the-art models, achieving 97.1% accuracy on the KUL dataset and 88.2% on the DTU dataset with a 2 s window, validated across group, subject-specific, and cross-subject analyses. Component studies confirm that integrating temporal and frequency features boosts performance, with the full TFDISNet surpassing its variants. Its dual-branch design and advanced loss functions establish a robust EEG-based AAD framework, setting a new field standard.