Abstract
Electroencephalography (EEG) is a widely-used noninvasive method to record brain activity. The high time resolution of EEG signals makes it a convenient method to analyze the time course of cortical activity during a prescribed task. Here, we focus on the task of reconstructing self-paced upper-limb movements directly from EEG signals. Numerous studies reported coherence between motoneuron activity and cortical activity in motor areas during motor tasks. Motoneuron activity relates to muscle force and thus to joint torque. Therefore, it is reasonable to presume that cortical activity reflects some aspects of joint kinematics during movement. However, EEG only captures a partial and filtered version of cortical activity. Our objective is to evaluate to what extent joint trajectories can be reconstructed from EEG signals.We address identifying features and locations of EEG signals that may reflect joint kinematics. First, we investigate measures to represent EEG signals. As EEG signals exhibit properties of nonlinear dynamical systems, we use complexity measures from chaos theory and statistical physics to complement classical spectral features in characterizing EEG signals. The quality of the complexity measures considered here depends on carefully selecting a resolution parameter. We propose a novel approach to determine this parameter that allows robust estimation of the measures. We validate our method on both simulated and real-world EEG data. Second, we evaluate the correlation of EEG complexity measures with EMG activity and joint kinematics. We recorded EEG, electromyography (EMG), and joint trajectories of 9 subjects performing self-paced cyclic elbow movements. By building statistical parametric maps of nonlinear EEG features, we identify the locations and features most correlated with movement.We then focus on reconstructing movement trajectories from EEG signals. In particular, we observe that the performance of models reconstructing motion trajectory from EEG signals is generally evaluated with correlation coefficients. We show that the test distribution of the coefficient degenerates for strongly correlated series such as movement trajectory. In that case, the test distribution of the correlation coefficient can be approximated by correcting its number of degrees of freedom. We propose a new parametric approach to estimate the number of degrees of freedom, which gives the appropriate test statistics. In light of the corrected distribution, we show the limited performances of models decoding movement trajectories from EEG signals. Finally, we construct a biologically-plausible model of the task involving a dynamic arm, a Hill-type muscle, a model of the spinal cord, and a neural mass model. We show perspectives and limitations arising from using a complex biophysical model with a state-of-the-art Bayesian inversion scheme.