Participants
Participants were a subgroup of patients selected from a larger clinical trial carried out at IRCCS Istituti Clinici Scientifici Maugeri, Lissone, Monza Brianza (MB), Italy (Clinical trial Identifier: NCT02439515). This clinical trial was aimed at investigating the effects of a multimodal biofeedback training, including FES-cycling, after stroke [27, 28]. Inclusion criteria were: first ever stroke within 6 months; hemiparesis secondary to a single unilateral stroke; low spasticity at the lower limbs (Modified Ashworth Scale < 2); ability to tolerate FES. Exclusion criteria were: allergy to stimulation electrodes; limitation at joint mobility; cognitive impairment (Mini Mental State Evaluation < 20); spatial hemineglect; other neurological comorbidities and presence of cardiac pacemakers.
Reference data obtained in a previous study [11] on a group of 12 healthy subjects (8 males, 4 females, mean age of 68 ± 5 years) were used for comparisons
The Ethical committee of the rehabilitation center approved the study (date of approval: 10/03/2014) and all subjects provided a written informed consent.
Intervention
Participants underwent a neuro-rehabilitative in-patient intervention for 3 weeks, 5 times per week. Each session consisted of 60 min of usual care and 25 min of voluntary cycling on a motorized cycle-ergometer (MOTOmed Viva2 ergometer, Reck GmbH, Germany) augmented by FES. The ergometer was equipped with force sensors (PowerForce™, Radlabor GmbH), which measured the radial and tangential forces at the two pedals. A visual feedback of the tangential forces produced by the two legs was provided to promote the execution of a symmetrical task [27, 28]. An 8-channel current-controlled stimulator (RehaMove2™ Hasomed GmbH, Germany) coordinated the bilateral neuromuscular stimulation of the quadriceps, hamstrings, tibialis anterior and gastrocnemius lateralis muscles, according to a biomimetic stimulation strategy, defined on the basis of physiological muscle activation strategy of a group of young healthy subjects while voluntarily pedaling on the same ergometer [32]. Self-adhesive bipolar surface electrodes (Pals® electrodes, Axelgaard Manufacturing Co., Ltd.), applied proximally and distally to the motor point of the muscles, delivered rectangular biphasic pulses with a pulse width of 400 μs and a stimulation frequency of 20 Hz. Current amplitude on each muscle was customized for each patient and could vary among participants in accordance to individual comfort. The current amplitudes delivered to the affected leg was chosen to produce a visible tolerated functional muscle contraction, while the unaffected leg received a stimulation intensity just above the sensory threshold. Pedaling resistance was set at a comfortable level that still permitted participants to smoothly cycle for the entire training session. Figure 1 shows the experimental setup used for training, the FES electrodes placement, as well as the intervention schedule.
Assessment
Participants were tested through clinical scales, gait analysis and neuro-mechanical evaluation of voluntary pedaling before (T1) and after (T2) the end of the intervention.
The lower limb component of the Motricity Index (range: 0–100) was used to grade the muscle power of the affected lower limb. The Berg Balance Scale (BBS) and the motor subscale of the Functional Independent Measure (FIM mot) were adopted to determine the functional mobility (range: 0–56) and the required degree of assistance (range: 13–91), respectively. Higher values result in lower level of impairment.
Gait analysis was performed using the GaitRite mat (CIR System Inc. USA) to assess spatial and temporal gait parameters: gait speed (cm/s), step length (cm), step time (s), double support time (s), swing velocity (cm/s). Each participant completed 3 walking trials at comfortable speed and the results were averaged across trials.
All patients underwent a neuro-mechanical assessment of voluntary pedaling using the same sensorized cycle-ergometer used for training. The test consisted of four trials at different cadences, performed in a random order (20, 30, 40, 50 revolution per minute, RPM). Each trial started with 1 minute of passive cycling in which the ergometer’s motor maintained a constant cadence without any voluntary contribution of the patient (passive phase); then the cadence maintained by the motor was reduced by 10 RPM and the patient was asked to voluntarily pedal for 2 min maintaining the same target cadence of the passive phase (active phase) [11]. It was chosen to perform isokinetic trials to have more repeatable pedaling cycles in terms of forces and EMG activations. The motor was kept ON also during the active phase in order to guarantee a smooth pedaling movement; however, being the cadence maintained by the motor lower than the target cadence, the voluntary involvement of the patient was assured. A minimum wash-out period of 5 min was guaranteed between consecutive trials to prevent muscle fatigue.
During the trials, the crank angle and the tangential and radial forces at the two pedals were acquired with a sampling frequency of 1000 Hz. Surface EMG signals of 9 muscles of both lower limbs (Gluteus Maximus (Gmax), Biceps Femoris long head (BFlh), Biceps Femoris short head (BFsh), Gastrocnemius Medialis (GAS), Soleus (SOL), Tensor Fasciae Latae (TFL), Rectus Femoris (RF), Vastus Lateralis (VL), and Tibialis Anterior (TA)) were acquired by a multi-channel signal amplifier (Porti 32™, TMS International) and sampled at 1024 Hz. The position of the self-adhesive Ag/AgCl electrodes (Kendall™, COVIDIEN) and the preparation of the skin followed the Surface Electromyography for Non-Invasive Assessment of Muscles (SENIAM) recommendations [33].
Data processing
All signals recorded during the neuro-mechanical assessment were processed and synchronized using a customized MATLAB program (R2018b, MathWorks Inc., Natick, MA, USA). Trials at different cadences were analyzed separately.
Muscle synergies analysis
The raw EMG signals were firstly filtered with a 3rd order band-pass (20-400 Hz) Butterworth filter. Then, the filtered signals were full-wave rectified and low-pass filtered at 5 Hz to obtain the linear EMG envelope. These envelopes were segmented by pedaling cycle using the synchronized crank angle and only cycles of the active phase which matched the trial target cadence with a tolerable range of ± 4 RPM were considered for the subsequent analysis. The segmented EMG envelopes were resampled in 360-points vectors with a cubic spline approximation and the left-side profiles were shifted by 180° [11]. Finally, the EMG envelopes were normalized to the median peak value calculated across pedaling cycles within each trial.
The Weighted Nonnegative Matrix Factorization algorithm (WNMF) was used to extract muscle synergies [22]. The WNMF algorithm was selected to adjust the analysis in the cases in which some muscles were missing (mostly occurring at Gmax). Indeed, a weight of 1 was assigned to acquired muscles and of 0 to non-acquired muscles. The algorithm used a multiplicative update rules to iteratively decompose the EMG envelopes in which the minimum threshold for convergence was set at 1 × 10− 5 for at least 30 repetitions and the maximum number of repetitions was fixed at 1 × 103. The WNMF was applied 30 times on 30 different datasets in order to ensure extraction consistency. Each dataset was composed by the normalized EMG envelopes of 30 pedaling cycles randomly selected among the available ones. Therefore, the WNMF was applied on a matrix M(9 × 10800) where 9 are the recorded muscles and 10,800 are the total number of EMG samples used in the analysis (30 cycles × 360 samples per cycle). The median weights among the 30 extractions were computed and the dataset characterized by the highest cosine-similarity (normalized scalar product) with the median weights was chosen as the most representative one. To test the consistency of the median extraction, the mean value of the similarity between the weights of the selected dataset and the remaining 29 extractions was computed for each subject, trial, and synergy. Synergy extraction was applied varying the number of synergies S from 1 to 6 in order to evaluate the coordination complexity. The optimal choice of S was quantified as the smallest number which allowed to reconstruct the EMG envelopes with a total Variance Accounted For (VAF) higher than 90% [11, 22]. VAF values were computed as:
$$ {\mathrm{VAF}}_{\mathrm{S}}=1-\frac{\sum_{\mathrm{i}=1}^9{\sum}_{\mathrm{j}=1}^{10800}{\left({\mathrm{M}}_{\mathrm{i}\mathrm{j}}-\left({\mathrm{W}}_{\mathrm{i}\mathrm{S}}{\mathrm{H}}_{\mathrm{S}\mathrm{j}}\right)\right)}^2}{\sum_{\mathrm{i}=1}^9{\sum}_{\mathrm{j}=1}^{10800}{\left({\mathrm{M}}_{\mathrm{i}\mathrm{j}}\right)}^2} $$
where i and j respectively indicates one of the nine muscle and a sample of the EMG envelope, while M, W and H respectively indicates the normalized EMG envelopes, the muscle synergies weights, and the activation coefficients.
The total Variance Accounted For by the first synergy (VAF1) was used as an additional estimate of synergy complexity [22, 34].
Regardless the inter-subject differences in the number of extracted synergies S, we fixed S to four in order to compare with the behavior of the healthy subjects group. The average set of synergies weights WHEALTHY, obtained in a previous study [11], was used for comparison. The four muscle synergies were sorted following the order of healthy muscle synergies by applying cosine-similarity analysis. Each extracted muscle synergy was compared with the healthy weight vectors and the highest cosine-similarity with healthy weights was chosen as criterion to define the synergy order. To identify longitudinal changes in synergy composition, the ordered muscle weights extracted from the patients before and after the intervention were compared with WHEALTHY by means of the cosine-similarity.
Afterwards, the Nonnegative Matrix Reconstruction (NMR) algorithm [35] was applied, updating iteratively the activation coefficients and fixing the synergies weights for each iteration. The goodness of the reconstruction was assessed computing the VAF values of the NMR. VAF values higher than 0.75 indicate acceptable reconstructions [15]. WHEALTHY was firstly chosen as fixed weights to allow comparisons between stroke and healthy synergy activation profiles [11]. The NMR was then applied to the EMG envelopes at both evaluations (i.e., T1, T2) with the fixed synergy weights of pre-intervention to evaluate longitudinal changes of the activation profiles. The Shape Symmetry Index (SSI) [36] was chosen as the metric to describe the similarity between two reconstructed activation profiles after phase correction:
$$ {\mathrm{SSI}}_{\mathrm{j},\mathrm{l}}=\frac{{\mathrm{C}}_{{\mathrm{h}}_{\mathrm{j},\mathrm{l}}^{\mathrm{T}1}{\mathrm{h}}_{\mathrm{j},\mathrm{l}}^{\mathrm{T}2}}}{\sqrt{\sum_{\mathrm{i}=1}^{360}{\mathrm{h}}_{\mathrm{j},\mathrm{l}}^{\mathrm{T}1}\left(\mathrm{i}\right){\sum}_{\mathrm{i}=1}^{360}{\mathrm{h}}_{\mathrm{j},\mathrm{l}}^{\mathrm{T}2}\left(\mathrm{i}\right)}} $$
where hj, l indicates the reconstructed activation profile of the j-th muscle synergy (1 to 4) for the lower limb l (left of right), computed by fixing the muscle synergies of the pre-intervention (T1). \( {C}_{h_{j,l}^{T1}{h}_{j,l}^{T2}} \) represents the circular cross-correlation function at lag 0. This index could range from − 1 to 1, with higher values indicating more similar shape profile despite possible differences in amplitude.
Biomechanical metrics
The biomechanical metrics of the pedaling trials were derived from the force signals acquired in the same 30 cadence-matched pedaling cycles selected for the muscle synergies analysis. Crank angle, tangential and radial force signals were low-pass filtered at 10 Hz with a 3rd order Butterworth filter. As for EMG signals, forces were segmented by pedaling cycle using the synchronized crank angle and then resampled, for each cycle, on a 360-point vector as a function of the crank angle by means of a cubic spline approximation and the left-side profiles were shifted by 180°. The net mechanical work of each leg was computed as the integral of the tangential active force profile, which was derived as the difference between the mean tangential force computed during the active phase and the one obtained during the passive phase [11].
The index of mechanical effectiveness (IE) was used as metric to assess the work dissipation between the voluntary contribution of the tangential force and the total force, computed as the vector sum of the tangential and radial force [11]. Finally, the Area Symmetry Index (ASI) was computed firstly to compare the tangential active force profile of the affected and unaffected leg and then to analyze a single side at once (affected or unaffected) with respect to the mean force profile of the healthy control group [37].
Statistical analysis
The Wilcoxon signed-rank test was used to compare the pre and post-intervention values for all outcome measures (clinical scales, gait parameters, neuro-mechanical metrics derived both from muscle synergy and force analysis).
For biomechanical metrics (work, IE, ASI) and VAF1, the Mann-Whitney U test was used to evaluate differences between patients and healthy subjects [11]. Pearson Chi-squared test for frequencies was used to compare the number of extracted synergies between patients and healthy subjects [11].
The Friedman Test was applied to evaluate the effect of cadence for work, IE, ASI, and VAF1. In case of significant changes, a post hoc analysis through a Wilcoxon signed-rank test was performed.
Finally, the Spearman’s correlation coefficient between neuro-mechanical metrics (work and VAF1) and clinical outcome measures, including also the gait speed, was computed. All statistical analyses were performed with IBM SPSS Statistics v.25.