CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings

Researchers introduce CORTEG, a framework that adapts pretrained scalp-EEG foundation models to intracranial ECoG recordings, enabling brain-computer interfaces to learn across patients with minimal calibration time. The approach demonstrates competitive or superior performance on finger trajectory and audio envelope decoding tasks while reducing per-patient training requirements to 10-30 minutes.

CORTEG represents a meaningful advancement in brain-computer interface technology by solving a persistent practical constraint: the scarcity of intracranial ECoG data per patient. Traditional BCI systems require extensive subject-specific training, limiting clinical deployment and scalability. This research demonstrates that knowledge learned from abundant scalp-EEG datasets transfers effectively to higher-fidelity intracranial recordings through careful architectural design, including an electrode-aware spatial adapter and dual-stream tokenization for different frequency bands.

The foundation model approach reflects a broader paradigm shift in neurotechnology, mirroring gains achieved across machine learning by pretraining on large datasets and fine-tuning for specific tasks. Prior BCI work remained compartmentalized by patient due to individual neuroanatomical differences; CORTEG's cross-patient learning capability addresses this fragmentation. The framework's ability to calibrate new patients in minutes rather than hours represents genuine clinical utility, reducing setup burden for medical practitioners and patients.

For the BCI and neurotechnology industry, this validates that foundation models extend beyond language and vision domains into neural signal processing. Companies developing clinical BCIs, neuroprosthetics, or brain-computer communication systems gain a reusable pretraining paradigm that accelerates development cycles and reduces data collection requirements. The work strengthens investment theses in neural interface companies by demonstrating technical pathways to faster clinical deployment.

The statistical significance observed primarily on the audio task suggests domain-specific transfer effectiveness varies, indicating future work should explore what signal characteristics enable robust cross-modality adaptation. Scaling this approach to larger ECoG datasets and additional decoding tasks remains the immediate research frontier.

• →CORTEG successfully transfers scalp-EEG foundation model knowledge to intracranial ECoG recordings, enabling cross-patient learning and competitive decoding performance.
• →New patients can be calibrated in 10-30 minutes on a single GPU, substantially reducing setup time compared to traditional subject-specific BCI approaches.
• →The framework achieved highest mean correlation on finger trajectory regression and statistically significant improvements on audio envelope regression tasks.
• →Cross-modality transfer effectiveness varies by task, with stronger results on audio decoding, suggesting domain-specific signal characteristics influence adaptation success.
• →Foundation model pretraining applied to neural signal processing opens new pathways for data-efficient clinical BCI deployment and patient accessibility.