EEG topographies provide subject-specific correlates of motor control

Electroencephalography (EEG) of brain activity can be represented in terms of dynamically changing topographies (microstates). Notably, spontaneous brain activity recorded at rest can be characterized by four distinctive topographies. Despite their well-established role during resting state, their implication in the generation of motor behavior is debated. Evidence of such a functional role of spontaneous brain activity would provide support for the design of novel and sensitive biomarkers in neurological disorders. Here we examined whether and to what extent intrinsic brain activity contributes and plays a functional role during natural motor behaviors. For this we first extracted subject-specific EEG microstates and muscle synergies during reaching-and-grasping movements in healthy volunteers. We show that, in every subject, well-known resting-state microstates persist during movement execution with similar topographies and temporal characteristics, but are supplemented by novel task-related microstates. We then show that the subject-specific microstates’ dynamical organization correlates with the activation of muscle synergies and can be used to decode individual grasping movements with high accuracy. These findings provide first evidence that spontaneous brain activity encodes detailed information about motor control, offering as such the prospect of a novel tool for the definition of subject-specific biomarkers of brain plasticity and recovery in neuro-motor disorders.

were chosen to mimic everyday life movements 1

Temporal characteristics of EEG microstates
To characterize and compare EEG microstates across different conditions, mean microstate duration, mean number of microstates per second, and percentage of the total analysis time covered by each microstate 2 were computed. Values were calculated for each subject and epoch. Subject-specific values were obtained averaging across epochs and were compared between conditions using Wilcoxon rank sum test (α=0.05) Bonferroni corrected for the number of comparisons (i.e., 15 for microstates A, B, C, and D, 10 for microstate E, and 6 for microstate F) 3 .

Statistical quantification of EEG microstate dynamics during resting state
We evaluated EEG microstate dynamics by computing EEG microstate occurrences. The latter was obtained by calculating the histogram of the most prevalent microstate within each temporal window (100ms) for each subject independently. In order to statistically quantify whether there was a modulation of EEG microstate dynamics over time, we divided the resting state period (i.e., one minute) in sliding windows of 2 seconds. We calculated significant differences between consecutive windows. For each comparison (i.e., comparison between consecutive windows), the significance threshold was obtained from a null-distribution constructed randomly permuting the microstates occurrence values of the two temporal windows compared. The number of permutations was determined to have α=0.05.

Additional LDA analysis
In the manuscript, we employed a Bayesian classifier, specifically a Linear Discriminant Analysis (LDA) 4 , to reveal the unique correspondence between microstates occurrences and motor task performed. Here we extended this analysis.
Specifically, we tested decoding accuracy i) when extracting microstates only in the training epochs, and ii) when using only movement execution phase as feature (in this case the dimension of the feature vector was 45, i.e., 9 time windows per 5 microstates). For the second test, we used the EEG microstates extracted averaging the signal over all epochs (i.e., microstates reported in the main manuscript).
For the first test, we selected a single repetition of the cross-validation procedure.
We extracted microstates for each grasp and subject independently for the averaged signals over the training epochs (i.e., half of the epochs of each grasp randomly chosen). The extracted microstates were then used to evaluate EEG microstate occurrences for both dataset (i.e., testing and training). The testing dataset corresponded to the remaining epochs not used to extract the microstates. For each subject and grasp type independently, a four-class LDA classifier was built using the microstates occurrences of the training epochs, and it was tested using the microstates occurrences of the testing epochs. As previously, the feature vectors used for the LDA classifier consisted of the microstate occurrences over movement preparation.

Temporal characteristics of the resting-state microstates are preserved
We assessed differences across conditions by evaluating the microstates temporal characteristics (i.e., mean duration, mean number of microstates, and total time covered, Figure S1).
The four resting-state EEG microstates showed a slightly lower mean duration than the ones of age-matched (25-30 years old) healthy subjects reported in literature 2 .
However, the mean number of microstates and the total time covered were comparable with previous results for all microstates 2 .
In general, during movement execution and holding phase, all the rest-specific microstates (i.e., A, B, C, and D) showed similar temporal characteristic compared to resting state. Only, microstate C, for pure-reaching movements and for power grasp (i.e., five-finger pinch and cylindrical grasp), and microstate D, for reaching-andgrasping movements, showed a reduced frequency of occurrence during the movement phase (Wilcoxon test, p<0.003). In addition, microstates A and C had a significant shorter total time covered for pure-reaching during the movement phase (Wilcoxon test, p<0.003).
No significant differences were found across motor tasks, grasp types, and between movement and holding phases, except for microstate C. In the manuscript, for each subject and dataset independently, subject-specific muscle synergies were extracted by utilizing the L2-norm NNMF algorithm 5 . Here, we checked consistency of results when using the KL divergence NNMF algorithm. Also for KL divergence we found that five (4.88±0.29 across subjects and motor tasks) and four (4.03± 0.42 across subjects and motor tasks) muscle synergies were sufficient to reconstruct more than 98% of the variance in the original signals respectively for movement execution and holding phase (see Figure S2c). The muscle weights of the synergies were highly similar between the two algorithms (mean DOT=0.98 and DOT=0.94 for movement execution and holding phase, respectively). Also the matching across motor tasks was preserved when using the KL algorithm. Indeed, synergies Syn 1, 2, and 3 were common across motor tasks (mean DOT=0.89). The fourth synergy (Syn 4), instead, was grasping-specific (mean DOT=0.97 across grasp types and mean DOT=0.44 between pure-reaching and reaching-and-grasping). Surprisingly the fifth synergy (Syn 5), which represented the contribution of the finger flexors and was not present in the holding phase except in pure-reaching, was substituted in five-finger pinch by an additional synergy (Syn 6) for the control of the thumb (mean DOT=0.65).

Microstates prediction of motor task is preserved when using only part of the data to extract them
In both tests (i.e., microstate extractions for training epochs and movement execution phase prediction), the decoding accuracy obtained (70% and 62% for the first and second test, respectively) was comparable to the one attained when extracting microstates in averaged signals containing both training and testing epochs during movement preparation.    Therefore, we can conclude that the presence of microstates was not modulated over time during resting state. Figure S5. Additional LDA results. We tested the decoding accuracy when extracting the EEG microstates using only the training epochs (a). Confusion matrices for the four grasp types were averaged over subjects. Grey levels ([0 100%]) code the decoding accuracy values. We then tested the decoding accuracy when using only movement execution phase as feature (b). Confusion matrices for the four grasp types were averaged over subjects and cross-validation repetitions.