Abstract
This paper proposes a novel simultaneous estimation method for EEG data recorded on different days. Our approach uses a generalized inverse filter that converts the acquired EEG data to cortical and eye currents. This method can quickly isolate cortical activities and eye artifacts from the EEG data regardless of the existence of sensor position changes.
Introduction
Electroencephalography (EEG) is a critical tool for Brain-Machine Interfaces (BMIs), yet its practical application is often hindered by the sensitivity of inverse filters to electrode placement. In multi-day studies, it is nearly impossible to position sensors in the exact same coordinates across sessions. Even minor shifts can lead to significant errors in estimating cortical activity. Standard recalibration methods to account for these shifts often require hours of data collection and computation, making daily BMI use cumbersome. This study addresses this challenge by proposing a method that decouples stable brain activity patterns from variable sensor geometry.
Methodology
The core of the proposed approach is a generalized inverse filter based on the Hierarchical Bayesian method (VBMEG). The researchers identified that while sensor positions (represented by the lead-field matrix, $G$) change between sessions, the underlying brain activity patterns (the current variance, $\Sigma_{\alpha}$) remain relatively stable for a given individual.
By reusing the variance calculated on an initial day and only updating the lead-field matrix with new sensor coordinates, a precise inverse filter can be generated in approximately five minutes. Furthermore, the model treats eye movements (EOG) as additional dipole sources rather than noise to be filtered out. This allows for the simultaneous estimation and separation of cortical and eye currents.
Data and Simulation
The method was validated using artificial EEG data constructed from high-fidelity human models. Structural data utilized a whole-brain model with 20,004 dipoles derived from a subject's MRI. Functional activity was simulated based on fMRI data from a smooth pursuit eye movement task.
To simulate real-world sensor displacement, 3D coordinates were recorded over three separate days using a FASTRAK system. The average sensor shifts were measured at 1.75 cm on Day 2 and 2.70 cm on Day 3 relative to the initial Day 1 positions.
Results
The proposed method successfully recovered cortical current estimation accuracy to levels comparable with "perfect" same-day measurements, effectively mitigating the impact of sensor shifts. Eye artifact suppression was highly effective, maintaining a correlation of approximately 0.99 between estimated and actual eye activity despite sensor displacement. Most significantly, the computation time for the new filter was reduced to approximately 5 minutes on a standard workstation, compared to the several hours typically required for a full hierarchical Bayesian estimation.
Conclusion
This research presents a viable path for making daily EEG-based BMIs practical. By treating sensor shifts as a geometric update and modeling artifacts as signal sources, the system achieves high performance with minimal recalibration time. Future work will focus on validating the method with real-world EEG data to confirm if the 5-minute recalibration holds up under varied noise conditions.
Summary updated