Cortical connectivity in people with Spinal Cord Injury during attempted arm and hand movements

Introduction: In this work, our main focus was to examine the time-varying (TV) cortical connectivity patterns that arise in SCI subjects during attempted arm/hand movements (i.e., supination, pronation, hand open, palmar/lateral grasp) [1]. To this end, we employed an adaptive formulation of the so-called multivariate autoregressive (TV-MVAR) models [2]. MVAR models provide metrics that describe causal interactions and directional effects between different signals. Herein, we used the TV directed coherence (DC) to quantify temporal variations in the direction of information transmission in the frequency domain. Specifically, we aimed to capture the general dynamics during various attempts of arm/hand movements and identify/localize the main sources of information flow.
Methods: 61-channel EEG signals were pre-processed (ICA-based artifact removal, trial rejection) using EEGLAB and Matlab. Source localization was carried out in Brainstorm (minimum norm imaging and sLoreta). Since our main scope was to examine connectivity related to motor function, we extracted 26 spatially segregated signals from anatomical regions defined by the Brodmann atlas (Fig.1a). For each type of attempted movement, we used the corresponding source signals to estimate a TV-MVAR model based on multiple trials from all subjects. To capture possible time variations, we applied a Kalman filtering approach. This approach assumes that the model coefficients are not constant but follow a random walk. Based on the estimated models we obtained DC time-frequency distributions. At each time point the total information outflow [3] from a particular region was defined as the sum of statistically significant connections (e.g., DC values) towards all other cortical regions.
Results and Discussion: In Fig.1b,c we present overlay plots depicting the total TV information outflow calculated from all attempted movements in the frequency range of [0.3 70] Hz, along with the average EEG signal in the sensor space. First, we observed an ipsilateral pattern (all subjects were right-handed), whereby dominant sources of information originated mainly from the right hemisphere. Second, the most prominent sources were detected in the sensorimotor and the primary motor area (specifically BA3a and BA4p) followed by the perirhinal and visual cortex (mainly in the contralateral side). Information outflow exhibited temporal patterns related to the onset of the cue. Sensorimotor outflow increased 0.5 sec prior to the cue and decreased after the onset of the cue. The opposite pattern was observed in the primary motor area, where outflow started increasing 0.5 sec after the cue onset and attained its maximum during the negative reflection of the average sensor signal (+1 sec). The visual and perirhinal cortex displayed increased outflow during the start of the trial and immediately after the cue onset, indicating possibly cognitive processing. This validates the hypothesis in [1], that the positive peak in the MRCP (+0.5s) is related with the presentation of the class cue. We also observed changes in the outflow in different frequency bands before and after the cue. Delta band outflow increased overall by 7.25% after cue onset, whereas the outflow in the rest of the bands decreased. The maximum percentage decrease was found in the beta band (7%). This could be attributed to possible event-related desynchronization phenomena.
Significance: The DC time-frequency distributions for different type of attempted movements followed approximately the same temporal trends. However, we detected discriminative connectivity patterns in different frequency bands and between different regions. Our future goal is to incorporate this information and improve BCI decoding performance [1] with respect to different movements.