Related Experiment Videos
Robust Averaging of Covariances for EEG Recordings Classification in Motor Imagery Brain-Computer Interfaces
Takashi Uehara1, Matteo Sartori2, Toshihisa Tanaka3
1Department of Electrical and Electronic Engineering, Tokyo University of Agriculture and Technology, Tokyo 184-8588, Japan.
Neural Computation
|April 15, 2017
Summary
Robust trimmed averages significantly improve electroencephalogram (EEG) classification accuracy in motor imagery-based brain-computer interfaces (MI-BCI) by effectively handling outliers in covariance matrix estimation.
Area of Science:
- Signal Processing
- Biomedical Engineering
- Machine Learning
Background:
- Covariance matrix estimation is critical for analyzing multivariate signals, particularly in motor imagery-based brain-computer interfaces (MI-BCI).
- Accurate covariance estimation is essential for feature extraction from electroencephalograms (EEGs) and subsequent EEG classification in MI-BCI.
- Tangent Space Mapping (TSM) is a key feature extraction method in MI-BCI that relies heavily on a well-chosen reference covariance matrix.
Purpose of the Study:
- To investigate robust algorithms for averaging sample covariance matrices (SCMs) to select a reliable reference matrix for TSM-based MI-BCI.
- To address the challenge of outliers in observed EEG signals that can negatively impact conventional covariance matrix averaging methods.
- To evaluate the effectiveness of robust estimators, specifically geometric medians and trimmed averages, in improving EEG classification accuracy.
Main Methods:
- Comparison of conventional geometric mean averaging with robust estimators like geometric medians and trimmed averages for SCMs.
- Implementation of trimmed averages by eliminating SCMs with the largest distances from the overall average covariance.
- Experimental validation using electroencephalographic recordings from MI-BCI subjects.
Main Results:
- Geometric medians demonstrated minimal improvement in classification accuracy compared to conventional methods.
- Trimmed averages, designed to mitigate outlier effects, yielded significant improvements in EEG classification accuracy across all subjects.
- The proposed trimmed average method effectively enhanced the performance of TSM-based MI-BCI.
Conclusions:
- Robust estimation techniques, particularly trimmed averages, are superior to conventional methods for selecting reference covariance matrices in TSM-based MI-BCI.
- Trimmed averages offer a practical solution for handling outliers in EEG data, leading to enhanced MI-BCI performance.
- Further research into robust statistical methods can significantly advance the field of brain-computer interfaces.