Participants
Fourteen able-bodied male individuals participated in this study (12 right-footed). All individuals provided written informed consent prior to participation in the experiment. The datasets from two participants had to be excluded due to poor quality of their EEG signals (after visual inspection and epoch rejection). Thus, data from 12 participants (21 ± 2 years, 83 ± 12 kg, 186 ± 6 cm) were analyzed. The experimental procedure was approved by the Ethical Committee for the department of Human Movement Sciences of the University of Groningen. The procedures complied with the guidelines defined in the Declaration of Helsinki [21].
Experimental procedure
Participants walked on a treadmill with and without a dummy mechanical knee prosthesis that simulates walking with a transfemoral prosthesis [3, 22, 23]. None of the participants had previous experience with the dummy prosthesis, and no specific instructions were given with respect to the use of the prosthesis or to the usage of the handrails during walking. The walking speed was kept constant across all participants at 0.9 m/s. The walking speed was selected based on energy efficiency and the average walking speed of people with a transfemoral amputation [24, 25]. All participants completed four blocks of 4-min treadmill walking with resting periods of 4 min in between (see Fig. 1).
In the first block, participants walked on the treadmill without the dummy prosthesis. After the first block, the dummy prosthesis was fitted to the right leg of the participants. Participants were not allowed to practice walking on the prosthesis before the measurement, but were allowed to flex, extend and support themselves with the knee of the dummy prosthesis, to familiarize themselves with the mechanism of the prosthesis. In the second, third, and fourth blocks, the participants used the dummy mechanical knee prosthesis to walk. Figure 1 illustrates the experimental setup and the timing of the procedure.
Data acquisition
Multi-channel electroencephalogram (EEG), ground reaction forces (GRF), and center of pressure (CoP) were recorded throughout the experiment. The EEG was recorded with 32 active Ag–AgCl electrodes (EasyCap GmbH, Herrsching, Germany) distributed across the scalp according to the international 10–20 system [26], using a wireless amplifier (Siesta, Compumedics Neuroscan, Australia) and the Profusion EEG software (Compumedics Neuroscan, Australia). The sampling rate was 512 Hz. Before each walking block, EEG was recorded for 2 min during quiet stance. To reduce potential artifacts in the EEG, the participants were instructed to limit their head movements and, to avoid talking, and excessive blinking.
The GRF and CoP for each foot were separately recorded with two force plates embedded in the treadmill (M-Gait, Motekforce Link, Netherlands). These data were recorded with D-Flow 3.26.0 (Motekforce Link, Netherlands) with a variable sampling frequency (later resampled at 300 Hz). A digital trigger was simultaneously recorded by both systems (Profusion and D-Flow) for synchronization.
Data analysis
All analyses were performed using MATLAB version 2014b (The MathWorks Inc., USA) with the addition of EEGLAB 14.1.2b (Swartz Center for Computational Neuroscience, USA) for EEG analyses.
Gait cycle segmentation
The gait events for heel strike and toe off were extracted from the GRF via threshold detection. The GRF data were filtered with a zero-lag low-pass 4th order Butterworth filter (10 Hz) and compared against a force threshold set to 30 N. Heel strike events were detected when the GRF exceeded the force threshold. Similarly, toe off events were detected when the GRF dropped below the force threshold. The gait events were aligned with the EEG using the digital trigger for synchronization.
Gait parameters
To assess any modifications of the gait pattern, the following gait parameters were computed: gait cycle duration, stance phase duration, step width, maximal GRF, and the GRF trace over time. The gait cycle duration was defined as the time difference between consecutive right heel strikes. The stance phase duration was defined as the percentage of the gait cycle spent between heel strike and toe off from the same foot. The step width was defined as the mediolateral distance of the filtered CoP (zero-lag, band-pass 4th order Butterworth filter, 0.5–15 Hz) between both feet during the double support phase of the gait cycle. The maximal GRF was defined through each gait cycle and for each foot, and it was normalized by the participant’s bodyweight (in Newton). The GRF trace was segmented according to the gait cycles of the right foot and time-normalized for gait cycle duration and the following fixed gait events: heel strike right (0%), toe off left (12%), heel strike left (50%), toe off right (62%) and heel strike right (100%). For group-level analyses, the mean of each parameter (gait cycle duration, stance phase duration, step width, maximal GRF, and the GRF trace over time) was computed (per participant) over all gait cycles within each walking condition.
EEG analysis
A schematic overview of the EEG processing steps can be found in Additional file 1. This approach is in line with previous studies on cortical dynamics during whole-body movement [15, 20, 27]. During acquisition, the EEG was filtered with a notch filter (50 Hz) to remove line noise. After acquisition, the EEG was filtered with a zero-phase high-pass FIR filter (1 Hz) and further processed with the CleanLine EEGLAB plugin [28, 29] to reduce line noise harmonics (100 and 150 Hz). The EEG was visually inspected for artifacts and noisy channels. Only one channel was removed in two of the participants.
The EEG was re-referenced to the common average and processed with the artifact subspace reconstruction (ASR) EEGLAB plugin that was used [30] to automatically remove non-stationary large-amplitude artifacts from the data. During a calibration stage, the ASR method determines a noise-free subspace from the continuous data, via principal component analysis (PCA). Then, a sliding-window PCA is computed over the data and compared against the noise-free subspace. If the variance of any principal component is above a certain threshold, the principal component is labeled as artifact and removed from the data. To ensure proper calibration of the ASR, the quiet stance EEG data were appended to the experimental data recorded during the walking blocks. The ASR user interface was configured to remove channels if the correlation with surrounding channels was less than 0.5, to reconstruct artifacts lying beyond ten standard deviations from the calibration data, and to remove a 500 ms time window from all channels if more than 25% of the channels contained artifacts at that moment in time.
After preprocessing, the EEG data were segmented into epochs ranging from − 0.4 to 2.2 s surrounding the right heel strike (i.e., the side of the dummy prosthesis). Epochs which did not contain a standard sequence of gait events (heel strike right, toe off left, heel strike left, toe off right, and heel strike right) were removed. Epochs with flat lines were visually identified and removed from the individual EEG datasets. The average number of remaining epochs (gait cycles) for the walking without dummy prosthesis, first, and last time walking with dummy prosthesis were (mean ± SD) 187 ± 13, 126 ± 26, and 146 ± 24, respectively.
Source separation
The segmented EEG data were separated into components from independent brain sources using Infomax independent component analyses (ICA) [31,32,33,34]. Then, the variance of individual epochs was computed for each independent component (IC) and normalized using the z-score per component across all epochs. Epochs with a normalized variance exceeding three standard deviations were marked as artifacts and removed from the data (resulting in 181 ± 14, 122 ± 26, and 141 ± 25 epochs remaining for walking without, first, and last time walking with dummy prosthesis). Afterwards the ICA was recomputed to ensure components were based on artifact-reduced EEG data. The resulting ICs were associated with an equivalent current dipole using a standardized three-shell boundary element head model (Montreal Neurological Institute (MNI)) and standard electrode positions (EEGLAB plugin DIPFIT; [35]). ICs were identified as possible brain sources according to their anatomical location (inside the head volume) and when residual variance of their equivalent current dipole was < 15% (mean number of ICs per participant: 3 ± 1.7, range 1–8).
The selected components were clustered across participants using the k-means clustering algorithm (k = 3) based on the following features: 3D anatomical location of their equivalent current dipoles, their mean power spectral density (PSD) (frequency band 3–48 Hz), their associated scalp projection, and their mean spectrogram across trials. These features were reduced to 10 principal components before clustering. Equivalent current dipoles (ECDs) which were located more than three times the standard deviation of distances within a cluster from any cluster centroid were considered outliers and were removed. Clusters with ECDs of at least half of the participants (n ≥ 6) were kept for statistical analysis. When clusters contained multiple ECDs of one participant, a single ECD with the shortest distance to the cluster centroid was retained for analysis. The Yale BioImage Suite [36] was used to determine the location of the cluster centroid and its corresponding Brodmann area.
Event-related spectral perturbation time–frequency maps
Event-related spectral perturbation (ERSP) time–frequency maps were used to compute modulations of intrinsic cortical rhythms [31, 37]. From each epoch (i.e., one gait cycle), single-trial spectrograms were computed and time-warped to normalize the duration of the gait cycle across all walking conditions.Footnote 1 The gait cycle and the gait events onset (i.e., heel strike right, toe off left, heel strike left, toe off right, and heel strike right) were normalized, using linear interpolation, to the median gait cycle duration and event onsets across all participants, all conditions, and all steps.
Average log-transformed spectrograms showing relative power changes were computed per individual IC and walking condition as the average difference between each (log-transformed) single-trial spectrogram and the average (log-transformed) spectrogram from the entire epoch (baseline). For visualization purposes, condition-specific baselines (i.e., the log-transformed power spectrum) were obtained from each walking condition over the complete gait cycle duration. Average time–frequency maps for a given IC cluster were computed by averaging across the maps corresponding to the ICs that were members of the cluster, separately for each walking condition.
Group-level statistical analyses
Gait parameters
Statistical analyses for the gait cycle duration, stance phase duration, step width, and maximal GRF were done with IBM SPSS Statistics 25 (IBM B.V., the Netherlands). The normal distribution of each parameter was first checked with the Shapiro–Wilk test. A repeated measures one-way ANOVA with post-hoc tests was conducted if the distribution was normal, otherwise a Friedman test was conducted with Wilcoxon signed rank tests for post-hoc testing. All post-hoc tests were Bonferroni corrected. The significance level for all tests was set at α = 0.05.
For the statistical analyzes of the GRF trace, MATLAB was used. A repeated measures one-way ANOVA was conducted of the normalized and time-warped GRF within the gait cycle. The significance level was set at α = 0.05 and it was corrected for false discovery rate [38] due to the multiple tests over the individual time points. Post-hoc comparisons were conducted with paired two-tailed t-tests and corrected in the same way.
Event-related spectral perturbation
MATLAB was used for comparison of the ERSP maps between conditions. The ERSP maps for the three walking conditions were computed with a common baseline (log-transformed spectrogram of walking without the dummy prosthesis). The significance of the modulations of cortical rhythms was determined with non-parametric permutation statistics [39, 40]. First, a one-way ANOVA of the ERSP maps with three levels (i.e., the walking conditions) was computed, and the resulting F-statistic per time point was stored. Then, a surrogate random distribution was created through random permutations of the condition labels (n = 200), followed by calculation of the surrogate F-statistic. The significance of the original F-statistic was determined by comparing against the surrogate distribution (critical alpha α = 0.05). Post-hoc tests (paired two-tailed t-tests) were conducted in a similar way. The significance level was corrected for false discovery rate [38].