CORTEG: Foundation Models Enable Cross-Modality Transfer from Scalp EEG to Intracranial Recordings
Researchers have introduced CORTEG, a novel cross-modality transfer framework designed to adapt large pretrained scalp-EEG foundation models for intracranial electrocorticography (ECoG) decoding. Addressing the challenge of limited per-patient data in brain-computer interfaces (BCIs), CORTEG leverages shared information across patients to enable efficient learning. The framework integrates a pretrained EEG backbone, an electrode-aware spatial adapter, and a dual-stream tokenizer, allowing calibration to new patients in just 10-30 minutes on a single GPU. Evaluations on finger trajectory and audio envelope regression tasks demonstrate that CORTEG matches or exceeds state-of-the-art baselines, with statistically significant improvements in low-data scenarios. This study provides systematic evidence that scalp-EEG pretraining can be effectively repurposed for ECoG, facilitating the development of data-efficient, adaptable intracranial BCIs. The findings highlight the potential of foundation models to bridge non-invasive and invasive neural recording modalities, enhancing decoding performance while reducing the data burden for individual patients.
Wire timeline
CORTEG: Foundation Models Enable Cross-Modality Transfer from Scalp EEG to Intracranial Recordings
Researchers have introduced CORTEG, a novel cross-modality transfer framework designed to adapt large pretrained scalp-EEG foundation models for intracranial electrocorticography (ECoG) decoding. Addressing the challenge of limited per-patient data in brain-computer interfaces (BCIs), CORTEG leverages shared information across patients to enable efficient learning. The framework integrates a pretrained EEG backbone, an electrode-aware spatial adapter, and a dual-stream tokenizer, allowing calibration to new patients in just 10-30 minutes on a single GPU. Evaluations on finger trajectory and audio envelope regression tasks demonstrate that CORTEG matches or exceeds state-of-the-art baselines, with statistically significant improvements in low-data scenarios. This study provides systematic evidence that scalp-EEG pretraining can be effectively repurposed for ECoG, facilitating the development of data-efficient, adaptable intracranial BCIs. The findings highlight the potential of foundation models to bridge non-invasive and invasive neural recording modalities, enhancing decoding performance while reducing the data burden for individual patients.
cs.AI updates on arXiv.org