CORTEG: Transferring Scalp EEG Foundation Models to Intracranial Recordings
A team of researchers has introduced CORTEG, a new framework designed for transferring data across different modalities. It adapts large pretrained scalp-EEG models for intracranial electrocorticography (ECoG) applications in brain-computer interfaces. This approach combines a pretrained EEG FM backbone with a specialized KNNSoftFourier spatial adapter, a dual-stream tokenizer for both low-frequency and high-gamma signals, and a fine-tuning method that leaves out one subject at a time. CORTEG allows for learning across different patients and can calibrate to a new patient in just 10-30 minutes using a single GPU. The framework was evaluated on two regression tasks, involving nine participants for finger trajectory and sixteen for audio envelope, achieving results that are on par with or better than earlier techniques.
Key facts
- CORTEG adapts scalp-EEG foundation models to ECoG
- Uses electrode-aware KNNSoftFourier spatial adapter
- Dual-stream tokenizer processes low-frequency and high-gamma activity
- Leave-one-subject-out fine-tuning strategy
- Calibrates to a new patient in 10-30 minutes on a single GPU
- Evaluated on finger trajectory regression (n=9) and audio envelope regression (n=16)
- Matches or exceeds prior decoding performance
- Enables cross-patient learning for brain-computer interfaces
Entities
—