Robust EEG-Based Motor Imagery Brain-Computer Interface using Lp-Norm-Based Common Spatial Patterns

Electroencephalogram (EEG) signals are inevitably contaminated by outliers, artifacts, and noise since EEGs are non-invasively recorded on the scalp. Motor imagery (MI) based brain-computer interface (BCI) makes use of multivariate EEG signals, which aims at classifying MI features. The existence of outliers leads to a degradation of the classification performance of MI-based BCI (MI-BCI). Previous popular common spatial patterns (CSPs) based on L2-norm and L1-norm dispersion are sensitive to outliers in EEG signals. This paper presents a generalized Lp-norm based CSP (p<1) to yield robustness to outliers for MI-BCI. By utilizing the Lp-norm dispersion of the filtered EEG samples instead of L2-norm or L1-norm ones, the robust MI-BCI to outliers is expected. Through simulations using a toy dataset and public EEG BCI dataset, we validate the capacity of the proposed Lp-norm BCI (CSP-Lp) and confirm its increased robustness to outliers.