Overcoming subject variability in P300 decoding based on quantification by deep learning

Binary P300 decoding is a cornerstone of P300-based brain-computer interfaces, enabling applications such as assistive communication, neurorehabilitation, and cognitive monitoring. However, substantial subject variability in EEG patterns poses a major challenge to the development of generalized P300 decoders. To address this problem, this study proposes SQ_EMoE, a subject-variability-aware enhanced Mixture-of-Experts framework that uses the confidence produced by a deep subject-identification network as a practical proxy for trial-level subject distinctiveness. High-confidence trials are routed to subject-specific experts, whereas uncertain trials are redirected to a pooled general expert, allowing the framework to dynamically balance subject-specific specialization and cross-subject generalization. Latent subject representations are further visualized to qualitatively examine subject separability and its relationship with routing behavior. Experiments on eight random subjects from each of two public P300 datasets yielded mean ROC-AUC values of approximately 0.86 on BNCI Horizon 2020 and 0.89 on P300-LINI, outperforming pooled EEGNet and architecture-matched pooled P300MCNN baselines. Ablation experiments further demonstrated the contribution of confidence-controlled fallback to the general expert. To evaluate scalability, the framework was additionally tested on all 22 subjects of P300-LINI. With the original subject-identification network, SQ_EMoE achieved a mean ROC-AUC of 0.8518; after strengthening the subject representation and identification network, performance increased to 0.8760, exceeding the pooled P300MCNN baseline (0.8482) and slightly surpassing the oracle routing or within-subject decoding (0.8744). Leave-one-subject-out experiments further showed that the under-realized SQ_EMoE conservatively redirects most completely unseen-subject trials to the general expert when reliable subject-specific routing is unavailable. These results demonstrate that effective modeling of subject variability, together with adaptive integration of shared and subject-specific representations, provides a promising alternative to conventional one-size-fits-all EEG decoding.