Zero-Shot Neural Priors Improve Generalizable EEG Decoding

Researchers have developed a novel framework aimed at improving the generalization of electroencephalography (EEG) decoding models. Traditional methods struggle with the high variability between individuals and the non-stationary nature of neural signals, limiting their applicability in real-world brain-computer interfaces (BCI) and mental health diagnostics. This new approach tackles these challenges by implementing a zero-shot cross-subject decoding framework. The framework leverages a Transformer-based foundation model, adapting it for regression tasks through a progressive unfreezing strategy. This technique helps prevent catastrophic forgetting while allowing the model to learn from a large dataset and apply that knowledge effectively to entirely new subjects without prior calibration. Benchmarked on the extensive Healthy Brain Network dataset, the fine-tuned Transformer model demonstrated superior performance compared to conventional baselines, achieving a notable improvement in normalized Root Mean Squared Error (nRMSE) on unseen subjects. This advancement paves the way for more scalable and calibration-free EEG decoding solutions, particularly beneficial for computational psychiatry and predicting behavioral outcomes.