Effect of power feature covariance shift on BCI spatial-filtering techniques: A comparative study.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Electroencephalogram (EEG)-based brain-computer interface (BCI) systems face performance degradation due to nonstationary input data distributions.
• Covariate shifts in signal features during intra-session transitions are a primary cause of this BCI performance compromise.

Purpose of the Study:

• To identify the most robust spatial filtering approach for EEG-based BCI systems under nonstationary conditions.
• To evaluate the effectiveness of Stationary Subspace Analysis (SSA) in mitigating performance decline and improving BCI accuracy.

Main Methods:

• Recorded calibration and test datasets from 17 healthy subjects for upper limb motor imagery BCI experiments, with a 30-minute interval between recordings.
• Applied and evaluated multiple spatial filtering methods (CSP, SPoC, SpecRCSP, SLap, SpecCSP, FBCSP, FBCSPT) on calibration data and tested their performance on the subsequent test dataset.
• Investigated the impact of Stationary Subspace Analysis (SSA) as a pre-processing step on model accuracy and the difference between calibration and test set performance.

Main Results:

• FBCSP and FBCSPT demonstrated superior robustness, maintaining median accuracy above 70% despite significant accuracy reductions observed with CSP, SPoC, and SpecRCSP.
• Stationary Subspace Analysis (SSA) pre-processing reduced the accuracy gap between calibration and test datasets and slightly increased overall accuracy.
• FBCSP and FBCSPT showed marginally better performance than other methods, even after SSA application.

Conclusions:

• Signal nonstationarity, driven by feature covariance shifts, significantly impacts BCI model accuracy, highlighting the need for robust evaluation frameworks simulating real-world performance.
• FBCSP and FBCSPT spatial filtering techniques are more resilient to feature covariance shifts, making them suitable for practical BCI applications.
• SSA effectively enhances BCI model performance and mitigates accuracy loss over time, improving reliability in dynamic environments.