Riemannian geometry-based detection of slow cortical potentials during movement preparation

Abstract

Slow cortical potentials - such as the readiness potential (RP) and the contingent negative variation (CNV) - provide means for controlling a brain-machine interface (BMI). RP and CNV precede self-paced and cued movements, respectively, and thus can be exploited for a variety of purposes such as robotic exoskeleton control or motor rehabilitation via BMI training. Single-trial detection of these patterns, however, is a challenging and therefore not yet fully resolved task, especially for online applications. Here we propose and evaluate a novel decoding algorithm for this cause, utilizing Riemannian geometry, template matching and adaptive re-centering for robust performance. We recruited 12 young, healthy volunteers who performed a center-out reaching task designed to evoke both RP and CNV in a sequential manner, while their neural activity was recorded using electroencephalography. Separate decoders of the same basic architecture were trained to detect RP and CNV on a single-trial basis, and data was evaluated offline on the subject level. RPs could be identified with a group average accuracy of 62.64±4.75% (ranging from 51.12% to 68.31%) with all but one subject surpassing chance levels, while CNVs were detected with an average accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{74.01\pm 7.49\%}$</tex> (ranging from 62.24% to 85.00%) with all participants surpassing chance level. Even though evaluation was carried out offline, the proposed pipeline is readily adaptable to an online setting. Therefore, our Riemannian geometry-based approach show potential in single-trial detection of slow cortical potentials.