Reproduce “IMU-Enhanced EEG Motion Artifact Removal with Fine-Tuned Large Brain Models” on OpenResearch

About this paper

Yuhong Zhang, Xusheng Zhu, Yuchen Xu, ChiaEn Lu, Hsinyu Shih, Gert Cauwenberghs, Tzyy-Ping Jung

Researchers at UCSD developed an IMU-enhanced, fine-tuned Large Brain Model (LaBraM) to address motion artifact contamination in EEG signals. Their method consistently reduced IMU–EEG correlation by over 50% compared to raw signals and over 13% against an ASR+ICA benchmark across diverse motion tasks and time scales.

Key insights

• The attention mechanism effectively aligned EEG and IMU features in a shared latent space, with learned attention weights closely mirroring frequency-domain correlation patterns, indicating accurate capture of cross-modal dependencies.
• Fine-tuning a pre-trained large brain model with a small fraction of domain-specific data (0.2346% of base model training data) can efficiently achieve substantial performance gains for specific downstream tasks like artifact removal.
• The model demonstrated interpretability by distinguishing between high and low correlation pairs between EEG and IMU, suggesting its ability to identify motion-related dependencies without explicit supervision.

Summary by alphaXiv.