Proceedings of the Annual Meeting of the Cognitive Science Society

The practical application of EEG emotion recognition is hindered by two critical challenges: reliance on offline processing pipelines that assume access to complete session data, and high inter-subject variability that complicates personalization. To address these, we propose BrainTune, an Artifact-Aware Meta-Contrastive Framework. BrainTune features a causal stream-compatible pipeline with an artifact-aware graph neural network to bypass offline dependencies, as well as a meta-contrastive strategy using supervised multi-level contrastive learning for the backbone to handle session/subject/task domains and meta-learning for the classifier to enable rapid personalization from brief calibration. Experiments on SEED, DEAP, and DREAMER demonstrate that BrainTune achieves state-of-the-art performance and exhibits "lasting personalization", where early calibration progressively benefits future sessions. Code is available at https://github.com/iewug/BrainTune.