Invasive neural recordings capture brain activity with high fidelity; intracranial EEG (iEEG) and intracortical spikes provide complementary spatial‑temporal information. Existing neural foundation models are mostly trained on a single modality, leaving joint pretraining on heterogeneous signals largely unexplored.
We introduce iBrain, which uses signal‑specific encoders for iEEG and spikes to respect their distinct characteristics, followed by a shared spatiotemporal Transformer backbone that models dependencies across channels and time. The model is pretrained on more than 7,000 hours of diverse recordings using masked signal reconstruction and channel‑view alignment, fostering contextual understanding and robustness across channels.
Results show iBrain consistently outperforms single‑signal baselines and reaches state‑of‑the‑art performance on several benchmarks. Further tests reveal strong transferability and data efficiency across different recording setups. These findings highlight the promise of joint pretraining on heterogeneous invasive recordings for scalable neural modeling and transferable representations.
Review