| name | corteg-eeg-ecog-cross-modality |
| description | CORTEG: Cross-modality transfer framework that adapts pretrained scalp-EEG foundation models to intracranial ECoG recordings. Combines EEG FM backbone with electrode-aware KNNSoftFourier spatial adapter, dual-stream tokenizer (low-frequency + high-gamma), and leave-one-subject-out fine-tuning. Enables competitive ECoG decoding with only 10-30 minutes of calibration data per patient. Activation: CORTEG, EEG foundation model, ECoG decoding, cross-modality transfer, scalp-to-intracranial, brain-computer interface, cross-patient learning, electrode-aware adapter. |
| arxiv_id | 2605.10337 |
| published | 2026-05-11 |
| authors | Liuyin Yang, Qiang Sun, Bob Van Dyck, Eva Calvo Merino, Marc M. Van Hulle |
| tags | ["eeg-foundation-models","ecog-decoding","cross-modality-transfer","brain-computer-interface","transfer-learning"] |
CORTEG: Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings
Foundation models pretrained on scalp EEG can be adapted for ECoG decoding, enabling cross-patient learning and competitive performance with minimal calibration data.
Source: arXiv: 2605.10337
Core Methodology
Key Innovation
The first framework to demonstrate that large pretrained scalp-EEG foundation models can be effectively transferred to the intracranial ECoG domain, bridging the gap between non-invasive and invasive recordings.
Technical Framework
- EEG Foundation Model Backbone: Leverage large pretrained models (e.g., from LAEEG, BENDR, or similar) that have learned generalizable features from scalp EEG data
- Electrode-Aware KNNSoftFourier Spatial Adapter: A novel spatial adapter that maps ECoG electrode positions to EEG foundation model input space, using k-nearest neighbor interpolation in the Fourier domain with soft assignment weights
- Dual-Stream Tokenizer: Separate processing streams for low-frequency (LFP-like, <100Hz) and high-gamma (>100Hz) activity, as these bands carry complementary information in ECoG
- Leave-One-Subject-Out Fine-Tuning: Train on N-1 patients, fine-tune minimally on held-out patient (10-30 min data, single GPU)
- Cross-Patient Decoding: Achieve competitive performance by sharing information across patients through the pretrained backbone
Key Results
- Successful cross-modality transfer: Scalp EEG foundation models adapt to ECoG with minimal fine-tuning
- 10-30 minute calibration: Competitive decoding performance on new patients with very limited data
- Dual-stream tokenization: Low-frequency + high-gamma streams outperform single-stream baselines
- Cross-patient learning: Information sharing across patients significantly outperforms patient-specific models trained from scratch
Implementation Patterns
CORTEG Framework:
1. Load pretrained EEG FM (weights frozen initially)
2. Add KNNSoftFourier spatial adapter for ECoG electrode layout
3. Add dual-stream tokenizer (LF path + HG path)
4. Train decoder head on N-1 subjects
5. Fine-tune spatial adapter + head on held-out subject (10-30 min)
Applications
- Clinical BCI: Rapid calibration for new patients using pre-existing EEG foundation models
- ECoG-based decoding: Motor imagery, speech decoding, cursor control
- Minimally-invasive interfaces: Bridge scalp EEG research to ECoG applications
- Cross-patient generalization: Leverage population-level knowledge for individual patients
Activation Keywords
- CORTEG framework
- EEG foundation model transfer
- ECoG cross-modality decoding
- KNNSoftFourier spatial adapter
- dual-stream brain tokenizer
- scalp-to-intracranial adaptation
- cross-patient ECoG learning
- brain-computer interface calibration
Related Skills
- neuroatlas-eeg-foundation-benchmark
- eeg-foundation-model-adapters
- nerve-network-aware-bilinear-fc-tokenization
- neural-encoding-evaluation-ground-truth
- what-do-eeg-foundation-models-capture
