AI Model Reconstructs Invasive Brain Signals from Scalp EEG
Researchers propose CAST, a machine learning framework that reconstructs intracranial EEG (iEEG) from non-invasive scalp EEG across different subjects, eliminating the need for patient-specific training data. The model uses a two-stage transfer learning strategy with a temporal encoder extracting multi-scale neural representations at three resolutions, followed by cross-attention mechanisms to map scalp signals to multi-channel iEEG waveforms. This approach addresses the circular dependency where prior models required invasive surgery to collect training data, limiting practical benefit. The study, published on arXiv (2605.18897), demonstrates cross-subject iEEG reconstruction for unseen patients, potentially advancing brain-computer interfaces and clinical applications without requiring surgery for each individual.
Key facts
- CAST (Cross-Attention Spatial-Temporal Transformer) reconstructs iEEG from scalp EEG
- Two-stage transfer learning strategy used
- Temporal encoder extracts multi-scale neural representations at three resolutions
- Cross-subject reconstruction eliminates need for patient-specific models
- Published on arXiv with ID 2605.18897
- Addresses circular dependency of prior patient-specific models
- Aims to benefit brain-computer interface and clinical applications
- Non-invasive scalp EEG used as input
Entities
Institutions
- arXiv