EEG-PRIME: Foundation Model for Cross-Dataset EEG Decoding.

Researchers have introduced EEG-PRIME, a novel two-stage foundation model specifically designed to overcome the significant challenges of generalizing Electroencephalography (EEG) decoding models across different datasets and individual subjects. These challenges typically arise from variations in acquisition protocols and unique neurophysiological characteristics. EEG-PRIME tackles this by integrating masked pretraining with a sophisticated prototype-aligned instruction tuning mechanism. During the pretraining phase, an EEG encoder learns highly transferable representations through masked reconstruction, enhanced by frequency-cutoff spectral augmentation. Subsequently, the instruction tuning stage incorporates multi-level conditioning, including task-semantic, dataset-specific, and crucially, subject-invariant signals. This conditioning modulates a Q-Former via Layer-wise Query Modulation, while frozen text embeddings of class labels serve as prototypes for cosine-similarity-based predictions across varied label spaces. Extensive experiments across sixteen diverse datasets covering various Brain-Computer Interface (BCI) paradigms demonstrate that EEG-PRIME consistently outperforms state-of-the-art baselines and previous EEG foundation models in cross-subject settings, even achieving zero-shot transfer capability on held-out datasets comparable to within-session calibration models.