Unified EEG Classification Across Subjects and BCI Paradigms Using Complementary Geometric and Spectral Representations

Deep learning is widely used for electroencephalography (EEG) decoding, but most models remain paradigm-specific. Unified decoding requires a single model to learn information across paradigms while retaining paradigm-dependent discriminative information. We propose the Complementary Dual-Expert Network (DuExNet), which is capable of 11-class cross-subject classification including motor imagery (MI), steady-state visual evoked potential (SSVEP), and mental workload (MW) paradigms. DuExNet consists of two complementary pathways. The first pathway learns a common global second-order representation enriched with spectral-statistical information. The second pathway captures sample-adaptive multiscale covariance structure to accommodate paradigm-dependent temporal variation. The calibrated probabilities from both pathways are combined by weighted fusion. We evaluated DuExNet using nine-fold cross-validation on three public datasets and compared it with DeepConvNet, EEGNet, CRGNet, LMDA-Net, and EEG-Deformer. DuExNet achieved an overall accuracy of 65.69% and a macro-averaged F1 score (Macro-F1) of 68.53%, exceeding the best counterpart by 4.25 and 4.55 percentage points, respectively. The statistical analysis showed that the performance improvement of the proposed model is significant. In addition, the ablation study demonstrated the effectiveness of combining the two pathways. Our study provides an effective model for cross-subject and cross-BCI-paradigm classification, aiming for universal models.