ModelsarXiv
NeurDuo-EEG: Long-Sequence EEG Foundation Model with Persistent State and Explicit Memory Released
NeurDuo-EEG introduces a new approach to EEG modeling with persistent, multi-timescale memory, significantly improving performance across short and long EEG analysis tasks. The model is open source and supports efficient streaming inference.
What Changed
NeurDuo-EEG is a new open-source EEG foundation model that introduces a novel multi-timescale memory management architecture. This enables the model to persistently model continuous EEG data with fixed-size state, overcoming the limitations of traditional EEG models which process data in fixed, independent windows.
Key Capabilities
- Causal EEG modeling with channel-resolved persistent memory.
- Multi-timescale memory management with learned consolidation and selective retrieval.
- Efficient streaming inference with nearly constant per-chunk latency for history up to one hour.
- Pre-trained on 3,955 hours of EEG from 17 public datasets using multichannel autoregressive prediction of discrete spectral codes.
- Achieves state-of-the-art results on four out of five EEG benchmarks, including substantial improvements in seizure detection.
What Developers Should Do
- Review the official code and documentation available at https://github.com/YifaNNW/NeurDuo-EEG.
- Evaluate replacing or augmenting fixed-window EEG processing pipelines with NeurDuo-EEG for improved long-duration and short-window analysis.
