Participants
Twelve stroke participants were included in the study (EXP group: 6 females/6males age 52.5 ± 18.5 yr) according to the following inclusion criteria: (1) first-ever unilateral, cortical, subcortical, or mixed stroke, caused by ischemia or hemorrhage (confirmed by magnetic resonance imaging), that occurred 3 to 12 months prior to study inclusion; (2) upper limb hemiparesis that was caused by the stroke; and (3) age between 18 and 80 years. The exclusion criteria were the presence of: (i) neuropsychological deficits preventing the ability to understand the instructions related to the experiment; (ii) concomitant diseases affecting the upper limb motor function (i.e., orthopedic injuries or other neurologic diseases affecting reaching or grasping); (iii) spasticity of each segment of the upper limb scored higher than 4 on the Modified Ashworth Scale (MAS [25]). All stroke participants were recruited within the inpatients and outpatients services of Fondazione Santa Lucia, IRCCS, Rome, Italy and were undergoing a rehabilitative treatment (usual care).
Twelve healthy participants (CTRL group: 9 females/3 males, age 43.6 ± 15.3 yr) participated in the study as a control group. Subjects did not present any evidence/known history of neurologic or neuromuscular disorders, nor any permanent/transient condition that could affect upper limb motor function.
The study was approved by the local ethics board at Fondazione Santa Lucia, IRCCS, Rome, Italy (CE PROG.752/2019) and all the participants signed an informed consent.
Clinical and functional evaluation was performed by expert physiotherapists before data acquisition (same day) by means of the following scales: (i) the National Institute of Health Stroke Scale (NIHSS) to assess general impairment derived from stroke [26]; (ii) the Manual Muscle Test (MMT) to assess strength in the paretic upper limb testing shoulder abduction, elbow flexion/extension and wrist flexion/extension [27]; (iii) the MAS scale to assess spasticity of shoulder, elbow and wrist muscles [25]. The upper extremity section of the Fugl-Meyer Assessment scale (FMA), comprising the four sub-scales “Upper Limb”, “Wrist”, “Hand”, “Coordination and Velocity” was performed to extensively describe the paretic upper limb residual function [28]. Handedness was assessed in all participants by means of the short form of the Edinburgh Handedness Inventory (EHI [29]).
Experimental design and data acquisition
The EEG and EMG signals were acquired simultaneously and sampled at 1 and 2 kHz, respectively. 61-channel EEG was recorded from the scalp by means of active electrodes (Brain Products GmbH, Germany) arranged according to an extension of 10–20 International System (reference on left mastoid and ground on right mastoid). Surface bipolar EMG signals were recorded by means of Pico EMG sensors (Cometa S.r.l., Italy) from the following 16 muscles: extensor digitorum (ED), flexor digitorum superficialis (FD), lateral head of the triceps muscle (TRI), long head of the biceps brachii muscle (BIC), pectoralis major (PEC), lateral deltoid (Lat_DELT), anterior deltoid (Ant_DELT) and upper trapezius (TRAP) of both sides (L: left, R: right for healthy subjects, AH: affected hand, UH: unaffected hand for stroke participants). EEG and EMG signals were amplified by means of BrainAmp (Brain Products GmbH, Germany) and Wave plus 16 channels (Cometa S.r.l., Italy) amplifiers, respectively.
The experimental setting is illustrated in Fig. 1. All participants were seated in a comfortable chair or wheelchair if needed, with their forearms resting on a pillow placed over a table (Fig. 1a). Participants were presented with visual cues displayed on a screen (1 m distance). The experimental session consisted of 4 runs (intermingled with breaks adapted to the patients’ necessities) during which the participant was asked to perform finger extension (Ext) and grasping (Grasp) with the right and the left hand separately (UH, AH for stroke participants). Each run comprised 40 trials (20 “task” trials of 8 s each and 20 “rest” trials of 4 s each in random order). The inter-trial-interval lasted 3 s during which participants were required to fixate a cross in the middle of the screen. “Task” trials started with 4 s of preparation (”get ready” instruction) afterward a go stimulus appeared (”task” instruction) and the participant had to perform the task for 4 s (Fig. 1b). In “rest” trials participants had to relax for 4 s (“relax” instruction—Fig. 1c). Participants were instructed to perform the task as fast as they could and to hold it at 15% of Maximum Voluntary Contraction (MVC) of the target muscle until the end of the trial (the experimenter guided the participants via online visualization of EMG traces). MVCs were recorded for each muscle at the beginning of the experiment for 5 s. Stroke participants attempted the movements with their affected limb to the best of their own residual ability, following the same instructions.
Data analysis
EEG–EMG data pre-processing
EEG data were band-pass filtered [3–60] Hz and Independent Component Analysis was used to remove ocular artifacts (Vision Analyzer 1.05 software, Brain Products GmbH, Gilching, Germany). EMG signals were downsampled to 1000 Hz, band-pass filtered [3–500] Hz and the electrocardiographic (ECG) component was rejected through template matching approach. A notch filter at 50 Hz was applied to remove power-line artifacts on both EEG and EMG signals. Task trials were segmented in 8 s epochs while Rest trials were segmented in 4 s epochs, both from the cue onset. To obtain EEG and EMG artifact-free trials, we applied a semi-automatic procedure. Specifically, for the EEG trials we defined a voltage threshold (∓ 100 μV) and rejected all trials in which 5 channels exceeded the threshold, otherwise a spherical interpolation was performed to replace noisy channels and the trial was saved. As for the EMG trials, we applied a statistical criterion based on the comparison between the EMG characteristics of each trial and the median EMG characteristics of all trials (reference characteristic) [30] then the selected trials were visually inspected and validated.
As for the EXP group, the EEG time series recorded over different scalp positions from patients with right-sided lesions were flipped along the midsagittal plane so that the ipsilesional side was common to all patients. Similar procedure was also applied to EMG data in all the patients with left affected hand (right hemisphere lesion). Both flipping procedures thus ensured to label the left hemisphere and contralateral right hand as “affected” in all the patients, independently from their actual lesion side.
Corticomuscular coherence (CMC) pattern computation
The EEG signals were re-referenced according to the common average reference (CAR) to correctly localize CMC peaks over sensorimotor areas in agreement with physiology of movement, as it has been demonstrated elsewhere [31]. The EEG edge electrodes were excluded from the analysis due to the possible presence of artifacts related to facial movements, thus only 41 EEG electrodes were included in the analysis. EMG signals were rectified before entering the coherence computation.
Coherence is an indicator of the linear connection between two signals, and it is an extension of Pearson correlation coefficient in the frequency domain. It is defined as cross-spectra normalized by auto-spectra [32]:
$$Co{h}_{xy}\left({f}_{j}\right)=\frac{{\left|{S}_{xy}\left({f}_{j}\right)\right|}^{2}}{\left|{S}_{xx}\left({f}_{j}\right)\right|\cdot |{S}_{yy}\left({f}_{j}\right)|}$$
(1)
where \({S}_{xy}\left({f}_{j}\right)\) is the cross-spectrum of signal x and y, while \({S}_{xx}\left({f}_{j}\right)\) and \({S}_{yy}\left({f}_{j}\right)\) are the power spectral densities of x and y respectively at a given frequency \({f}_{j}\).
Typically to test whether CMC is significant, its values are compared to the chance level. However, in motor tasks it is mandatory to go beyond the null-case validation and thus, to assess the significance of the connections against rest condition to ensure that only the relationships related to the executed task are kept. Accordingly, we decided to use a non-normalized version of the CMC in (1) to prevent the detection of false positives in CMC when the muscle activation level is around 0, as expected in the rest time interval of our experiment [33].
The CMC was computed in a 2 s-window which were selected differently for Task and Rest condition. As for “task” trials, the interval of [5–7] s from cue onset was selected whereas we selected the first artifact-free interval of 2 s length in “rest” trials.
CMC values were estimated in the range [1–60] Hz for each participant, movement (ExtR/AH, ExtL/UH, GraspR/AH, or GraspL/UH) and interval of interest (Task, Rest). Two different procedures were followed for the CMC estimation: across-trials and single-trials for Group Analysis and Single Subject Analysis, respectively. As for the across-trials approach (periodogram window length of 1 s with 0% overlap) a single CMC pattern was estimated from all trials in the dataset of a single participant, in order to have an average of CMC pattern for each single participant to enter in the grand average (see Statistical analysis—GA patterns). As for the single-trial approach (periodogram window length of 0.250 s with 50% overlap), a CMC spectrum was estimated for each trial in the dataset, to obtain different observations of CMC patterns for each single participant. The CMC values were then extracted for the 3 considered frequency bands defined as alpha (8–12 Hz), beta (13–30 Hz) and gamma (31–60 Hz). For each of these bands, we identified the characteristic frequency as the frequency in which CMC showed the highest value for each pair of signals. The characteristic frequency was specific for each pair of signals, it was computed in the Task condition and used also for the Rest in order to compare patterns at the same frequency.
Analysis of CMC patterns properties by graph theory indices
CMC networks estimated at single-subject level were assessed against chance level and thus transformed into weighted CMC adjacency matrices. The single-subject CMC adjacency matrices were built as follows: for each EEG–EMG pair we applied an unpaired t-test between task and rest conditions on CMC values estimated by means the single-trial procedure. The significance level was set to 0.05. False Discovery Rate (FDR) was used to control family-wise error rate [34]. Such statistical comparison was used to assess CMC values obtained during movement execution/attempt against chance level using as null-case statistical threshold the corresponding CMC values in rest condition. The application of this test allowed to obtain for each subject and each movement a CMC adjacency matrix where null-values correspond to EEG–EMG connections not significantly different from rest while non-null values correspond to connections where CMC values were significantly higher during movement than rest condition. The comparison between task and rest conditions allowed also to reduce the presence of spurious connections in CMC networks due to volume conduction which is an intrinsic phenomenon of the EEG signals.
The Graph Theory was applied to the obtained CMC adjacent matrices to extract a set of ad hoc indices which synthetically described the main properties of the CMC patterns. This procedure aimed at reducing the CMC matrix complexity and thus allowing its interpretation.
Such computation was repeated for each subject, movement, and band.
Global network properties:
-
CMC Weight is defined as the average of CMC values of the existing connections in the network. It is a measure of the strength of the EEG–EMG connections which is well-known to be reduced in stroke patients [13].
-
Network Density (ND) computed as the total number of existing connections in the pattern normalized for the possible number of connections.
Network density was also calculated for each of the identified 4 sub-networks as follows (local networks properties):
-
Density (of) Contralateral Hemisphere (DCH) calculated as the total number of existing connections that link the target muscle (FD in Grasp and ED in Ext) with EEG electrodes in contralateral hemisphere (normalized for the possible number of connections in this sub-network).
-
Density (of) Ipsilateral Hemisphere (DIH) calculated as the total number of existing connections that link the target muscle (FD in Grasp and ED in Ext) with EEG electrodes in ipsilateral hemisphere (normalized for the possible number of connections in this sub-network).
-
Density (of) Involved Side (DIS) calculated as the total number of existing connections entailing muscles in the side involved in a given motor task – target muscles (normalized for the possible number of connections in this sub-network).
-
Density (of) Uninvolved Side (DUS) calculated as the total number of existing connections entailing muscles in the side which is not involved in a given motor task – non-target muscles (normalized for the possible number of connections in this sub-network).
To further investigate the selective engagement of muscles, we computed:
-
Muscle Degree (MD) defined as the total number of connections that each muscle establishes with EEG channels normalized for the maximum number of possible connections involving it. This index allowed us to measure the involvement of each muscle in the pattern and to identify the muscles with a dominant role (higher degree) with respect to others. It was calculated for each of the recorded 16 muscles both during Ext and Grasp, and then a qualitative comparison was performed between the muscle degree values relative to the movement involved and uninvolved side.
-
Distal/Proximal Degree Ratio (DPDR) was computed considering the degree of the muscles of the movement involved side, that were labeled as distal (FD and ED) and proximal (BIC, TRI, Ant_DELT, Lat_DELT, PEC, TRAP). It was defined as the ratio between the degree of distal muscles and the sum of degrees in distal and proximal muscles. DPDR value was set as equal to: 1 if the activation regarded only distal muscles; 0 for the activation of only proximal muscles; 0.5 in the case of both proximal and distal muscle activation with the same weight.
Statistical analysis
Grand average (GA) CMC patterns
Each movement was described by a coherence pattern as a result of a GA analysis computed for the CMC values across participants (see Figs. 2 and 3). A paired sample t-test with the interval (Task vs Rest) as independent variable and the CMC values computed in the across-trials procedure as dependent variable was applied to each movement type, frequency band and channel pair. The significance level was set to 0.05. FDR was used to control family-wise error rate.
Between-groups differences in CMC pattern properties
A Kruskal–Wallis test was applied on each graph theory derived index considering as factor the three groups: CTRL—control group executing the task with the right hand; EXP_UH—stroke group executing the task with the unaffected hand; EXP_AH—stroke group executing the task with the affected hand. A Tukey’s post hoc test was applied to assess between groups differences. We selected the right hand for CTRL group since no significant differences were observed in the graph indices between left and right hand.
Correlation between brain network indices and functional/clinical scales
Brain network indices that significantly described the CMC patterns of stroke patients performing movements with the affected arm were correlated with the scores obtained from the following clinical scales: FMA total, FMA sub-scales and MMT. The Spearman’s correlation test was applied with the indices values as the dependent variable and the clinical scales’ scores as the independent variable.