Motor Imagery Brain-Computer Interfaces: Random Forests Vs Regularized Lda - Non-Linear Beats Linear
David Steyrl, Reinhold Scherer, Frstner, Oswin, Mller-Putz, Gernot R.
Abstract
Nowadays, non-linear classifiers are available that claim to generalize well at a low amount of data. Recently, we conducted an on-line study, where a random forest (RF) classifier successfully drove an electroencephalography (EEG) based sensorimotor rhythms (SMR) brain-computer interface (BCI) by classifying discrete Fourier transform (DFT) features. In this work, we re-analyse that data-set and simulate the use of common spatial patterns (CSP) features with a RF classifier and a shrinkage regularized linear discriminant analysis (sLDA). We found that the RF classifier could make better use of the CSP features and outperformed sLDA. The mean and median classification accuracy during the feedback period were improved by 2% and 3% when using a RF classifier. The eect is small, but statistically significant (p < 0.05) and consistent over the participants. Therefore, we argue that the widespread view that linear methods are ideal for BCIs should be reconsidered and RF classifiers should be taken into account when choosing a classifier for SMR-BCIs.