Proceedings of the Annual Meeting of the Cognitive Science Society

Driver vigilance estimation is critical for preventing fatigue-induced traffic accidents, yet existing multimodal EEG–EOG methods often suffer from limited personalization and poor generalization. We propose KANformer, a personalized vigilance estimation framework that integrates subject-specific priors, Transformer-based feature encoding, and Kolmogorov–Arnold Networks (KAN) for adaptive temporal modeling. Raw EEG and EOG signals are first encoded by a Transformer to capture long-range dependencies and cross-modal interactions. A personalized channel attention module then reweights multimodal features using demographic and task-related metadata, enabling subject-aware representation learning. These representations are further modeled by a KAN-based temporal module, which replaces Mamba-style state-space modeling with more expressive functional compositions while retaining low computational complexity. Experiments on the SEED-VIG and SADT datasets under cross-subject and zero-shot settings demonstrate that KANformer consistently outperforms Mamba-based baselines in RMSE, MAE, and PCC. The results indicate that KANformer provides a scalable and generalizable solution for real-time fatigue monitoring and personalized neurocognitive modeling.