Electrocortical theta activity may reflect sensory prediction errors during adaptation to a gradual gait perturbation

Participants

Thirty-three young, neurotypical adults (15 female, 18 male) participated in this study. All participants self-reported being right-foot dominant (to avoid brain lateralization confounds) and physically active (regularly exercising ≥ twice per week for ≥30 min) and had normal or corrected to normal vision. None of the participants had a history of major musculoskeletal, neurological, or cardiovascular conditions. Participants were excluded if they had previously walked on a split-belt treadmill because the gait task needed to be novel. The experimental protocol was approved by the University of Florida Institutional Review Board (IRB201701603) and follows to the standards set by the Declaration of Helsinki. All participants provided written informed consent before participating.

Experimental design and procedure

Figure 1 shows a diagram of the experimental setup. Participants were equipped with our dual-electrode EEG system (described in the subsequent section), neck electrodes for electromyography (EMG), and motion capture markers. While participants walked on the split-belt treadmill, we recorded ground reaction forces from force plates under each belt (Fs = 1,000 Hz) and lower body kinematics using motion capture (Fs = 100 Hz; Optitrack, Corvallis, OR, USA). These data were synchronized through the motion capture software’s eSync. We used the Rizzoli lower body marker set, which was comprised of 26 markers, along with four rigid bodies placed on the shanks and thighs. Participants also wore earplugs to reduce their ability to use the sound of the treadmill to identify differences in belt speeds, which was particularly important for the gradual adaptation condition. They were also told to refrain from looking downward at the treadmill belts and keeping track of their cadence by any means (e.g., singing songs in head, counting) to prevent them from identifying subtle speed changes throughout the adaptation conditions. A fixation cross was placed on the board in front of them. Participants were told that the purpose of this cross was to help them stay oriented and prevent lateral shifting on the treadmill and that they did not have to stare at it during the entirety of the experiment.

Our experiment had two adaptation conditions: abrupt and gradual. In both adaptation conditions, the final belt speed ratio was 2:1, with the belts moving at 1.2 m/s and 0.6 m/s under the right and left foot, respectively. We chose 1.2 m/s as our fast belt speed as it was close to the preferred treadmill walking speed of young adults (Terrier & Reynard, 2015) while likely keeping motion and muscle artifact contamination manageable. In the first adaptation condition, referred to as abrupt adaptation, we increased the error magnitude by abruptly perturbing the participants. During the second adaptation condition, referred to as gradual adaptation, we kept the error magnitude small by applying a gradual perturbation. We did not randomize the order of gradual and abrupt adaptation because this was a secondary analysis from a larger study that required the abrupt adaptation to be completely novel (Jacobsen & Ferris, 2023). However, results from Jacobsen & Ferris (2023) indicated that electrocortical (4–30 Hz) after-effects had washed out by the end of the first post-adaptation period.

Figure 2A shows the task paradigm. Participants walked on a split-belt treadmill (Bertec, Columbus, OH, USA) where each belt speed could be controlled independently. They were informed that the belt speeds would at some point be different, but not what the speeds would be. We told them when the treadmill was initially about to start or stop, but otherwise gave no indication of when speed changes would occur. We recorded all belt speed transition events with a button connected to the EEG system. To start, we recorded a pre-adaptation period in which participants walked with the belts moving together (i.e., tied) at three different speeds: slow (0.6 m/s), fast (1.2 m/s), and medium (0.9 m/s). During abrupt adaptation, participants encountered a sudden perturbation (a = 0.2 m/s²). The treadmill maintained a 2:1 belt speed ratio for the entire 15-min period. This was followed by a 10-minute post-adaptation period (post-adaptation #1) to wash-out after-effects, which is the minimum time suggested by Vasudevan, Hamzey & Kirk (2017). Roemmich & Bastian (2015) also used a 10-minute post-adaptation period to separate two different adaptation conditions (Roemmich & Bastian, 2015). Participants then experienced a gradual change in belt speeds. During the first five minutes of gradual adaptation, the right belt linearly increased the speed from 0.9 m/s to 1.2 m/s and the left belt linearly decreased the speed from 0.9 m/s to 0.6 m/s (a = 0.001 m/s²; similar acceleration to Long, Roemmich & Bastian, 2016). Once the belts reached terminal speed, the treadmill maintained a 2:1 ratio for two minutes. We limited the length of the gradual adaptation to keep entire experiment under 60 min; this was for the benefit of the volunteers and so that our results are comparable to clinical gait adaptation studies. Despite its brevity, we still expected motor adaptation to occur during gradual adaptation because after-effects can be present even after just two minutes of walking on a 2:1 split-belt treadmill (Roemmich & Bastian, 2015). The experiment ended with 5 min of tied-belt walking (referred to as post-adaptation #2) to wash out after-effects.

Dual-layer EEG and EMG setup

We used a custom dual-electrode system to conduct high-density EEG recordings. The system consisted of 128-pin type scalp electrodes and 128 flat-type “noise” electrodes using hardware from the BioSemi ActiveTwo system (Fs = 512 Hz; from Biosemi in Amsterdam, The Netherlands). For each scalp electrode, there was an inverted noise electrode physically attached (schematic in Fig. 3A). The noise electrodes were electrically isolated from the scalp electrodes, allowing movement artifacts and external noise to be exclusively recorded in separate channels. The design and functionality of the dual-electrode system was fully described in Nordin, Hairston & Ferris (2018); this previous work demonstrated a twofold enhancement in the signal-to-noise ratio through a phantom head simulation experiment.

We recorded neck EMG to isolate muscle activity and attenuate undesirable muscle signals in our EEG data. Richer et al. (2020) used a conductive phantom head and robotic motion platform to demonstrate the effectiveness of using neck EMG and dual-electrode EEG in improving recovery of complex artificial brain signals during motion (Richer et al., 2020). We used eight flat-type neck electrodes to record surface EMG signals from the left and right sternocleidomastoid, splenius capitis, and trapezius. The EMG electrodes had the same ground and reference as the scalp electrodes and were later rereferenced. We used a hard-wired 0.5 Hz square wave to synchronize the force plate, motion capture, EEG, noise, and EMG data. The data from the EMG and noise channels were used to attenuate artifacts in the EEG data, which we will elaborate on in the EEG Analysis section.

Multiple steps were taken prior to data recording to ensure robust EEG data were collected. Participants were instructed not to apply any hair products prior to the experiment. The placement of the EEG electrode cap followed the International 10–20 system. A reference line was marked on the forehead to monitor any potential shifting of the cap during the experiment. We recorded electrode locations using a 3D head scanner, with different scanners employed due to hardware and software issues (Eva scanner, Arctec3D, Santa Clara, CA, USA; Structure sensor on Apple^® iPad; Scanner, Occipital Inc., San Francisco, CA, USA; itSeez3D). To establish good contact between the electrodes and the scalp, a blunt-tip gel syringe was used to gently part the hair and lightly abrade the scalp before applying the gel and placing the electrodes. The scalp electrode offsets were adjusted to be at or below 20 mV prior to recording. Once the scalp and noise electrodes were gelled, we positioned a secondary custom conductive cap (Eeonyx, Pinole, CA) over the noise electrodes to mimic the scalp electrodes’ circuitry and conductivity. This second cap was secured with stretch self-adherent tape to prevent shifting, and all wires were tightly bundled with Velcro straps to minimize cable sway, which is known to introduce artifacts into EEG signals (Symeonidou et al., 2018). Prior to recording, participants were shown the effects of their own jaw clenching, excessive eye blinking, and neck muscle tension on the raw EEG data stream and were instructed to refrain from these behaviors.

Behavioral analysis

To identify gait events, we used vertical ground reaction force data from the instrumented treadmill. The ground reaction force data were low-pass filtered (zero-phase, second order, Butterworth filter) with a cutoff frequency of 6 Hz to remove noise and down sampled to match the EEG sampling rate (512 Hz). We used a 10 N threshold to detect foot contact and foot lift-off events. We visually inspected time windows with abnormal gait event latencies (>4 standard deviations (SD)) and corrected any missteps (such as stepping on the wrong force plate).

We binned strides across the stages of pre- and post-adaptation to create subconditions shown in Fig. 2B. Pre-abrupt and pre-gradual adaptation were defined as the 30 strides preceding adaptation. The post-adaptation #2 period, which begins when the belts abruptly return to 1:1 ratio, was divided into initial post-adaptation (strides 1–10), early post-adaptation (strides 11–40), and late post-adaptation (last 30 strides).

Step length asymmetry was our primary metric of kinematic error because it adapts robustly during split-belt treadmill walking (Reisman, Block & Bastian, 2005). To quantify step length, we measured the distance between two ankle markers in the anterior-posterior direction at the instant the leading leg foot made contact with the treadmill belt. This led to two separate measures of step length, one for each leading limb. Fast step length refers to the step length when the limb on the faster moving belt was leading, and vice versa for slow step length. We quantified step length asymmetry as the difference in step length between the fast limb, which is on the faster belt, and the slow limb, which is on the slower belt, normalized by the sum of the two step lengths following Reisman, Block & Bastian (2005): (1)${\displaystyle \text{step length asymmetry}=\frac{{\text{step length}}_{\text{fast}}-{\text{step length}}_{\text{slow}}}{{\text{step length}}_{\text{fast}}+{\text{step length}}_{\text{slow}}}.}$

We normalized step length asymmetry values by subtracting individual average step length asymmetry from the pre-abrupt adaptation condition (0.9 m/s tied-belt; i.e., within subject normalization) to compare behavior across participants.

We extracted asymmetric strides (i.e., kinematic errors) that occurred during adaptation so we could evaluate how electrocortical activity changes during stepping errors. We defined asymmetric strides as gait cycles with step length asymmetry values greater than two standard deviations from the mean of the last 30 strides of the reference condition preceding adaptation (i.e., pre-gradual adaptation): (2)${\displaystyle \left|SL{A}_{\text{asymmtric}}\right|>\left({\mu }_{\text{reference}}+2{\sigma }_{\text{reference}}\right)}$ where SLA is step length asymmetry and µ(σ) is the mean (standard deviation) of step length asymmetry from the reference condition. We used pre-gradual adaptation as the reference condition because in our previous study we found that behavioral after-effects during post-adaptation #1 did not fully washout (Jacobsen & Ferris, 2023). To determine the magnitude of the perturbation during the asymmetric strides, we calculated belt speeds using the average anterior-posterior velocity of the ankle markers during stance (when the foot was on the belt; Hoogkamer et al., 2015). We linearly interpolated gait variables and belt velocity missing values. Belt velocity data were smoothed using a finite impulse response filter (b = 1/6, a = 1). Belt velocity outliers were defined as elements more than three local standard deviations from the local mean within a six-element window. Outliers were then replaced with the mean of the strides preceding and following the outlier. We also calculated the belt symmetry in a similar way as step length asymmetry (following Hoogkamer et al., 2015): (3)${\displaystyle \text{belt symmetry}=\frac{{\text{belt speed}}_{\text{fast}}-{\text{belt speed}}_{\text{slow}}}{{\text{belt speed}}_{\text{fast}}+{\text{belt speed}}_{\text{slow}}}.}$

Because participants had completed two adaptation periods (abrupt and gradual), we tested whether participants who were good at adapting to an abrupt perturbation were also good at adapting to a gradual perturbation. We used step length asymmetry to assess behavioral performance between abrupt and gradual adaptation. To quantify abrupt adaptation performance, we calculated the number of strides it took to reach steady-state step length asymmetry (similar to Fettrow et al., 2021; Finley et al., 2015). Steady-state was defined as the average value during the last 30 steps of abrupt adaptation. The threshold for reaching steady-state was defined as the point where step length asymmetry remained within 2 standard deviations of the steady-state for at least 10 consecutive steps. To quantify gradual adaptation performance, we used a separate performance metric because asymmetric strides were not occurring in an isolated period that was easily distinguishable from steady-state. We measured gradual adaptation performance as the percentage of asymmetric strides. Both performance metrics were normalized between zero and one and inversed so that a value of one indicated fewer strides to reach steady-state step symmetry or lower percentage of asymmetric strides for abrupt and gradual adaptation, respectively (i.e., a value of one indicates better adaptation). We performed linear regression to investigate the relationship between abrupt and gradual adaptation performance. The α-level was set at P = 0.05.

Volume conduction modeling

We constructed a volume conduction model of each participant’s head using individual structural magnetic resonance images (T1-MRI) following the Fieldtrip-SIMBIO pipeline (Vorwerk et al., 2018; Vorwerk et al., 2018; Vorwerk et al., 2018) outlined in Fig. 3B. Previous studies have demonstrated that realistic head models improve the accuracy of the EEG inverse solution (Akalin Acar & Makeig, 2013; Haueisen et al., 1997; Neil Cuffin, 1996; Roth et al., 1993; Vanrumste et al., 2002). We used the finite element method to construct the head model and merged it with digitized electrode locations using fiducial markers. The custom head models consisted of five compartment layers with specific tissue conductivity values: skin = 0.43 S/m, skull = 0.01 S/m, cerebral spinal fluid = 1.79 S/m, gray matter = 0.33 S/m, white matter = 0.14 S/m (Baumann et al., 1997; Haueisen et al., 1997; Hoekema et al., 2003; Vorwerk et al., 2019).

EEG Analysis

The EEG processing pipeline is outlined in Fig. 3C. We processed data in Matlab 2022a (MathWorks, Natick, MA, USA) using EEGLAB (v2022.0) (Delorme & Makeig, 2004), Fieldtrip (v20210614) (Oostenveld et al., 2011), and custom scripts as detailed in Jacobsen & Ferris (2023). We used a combination of standard and novel pre-processing steps to clean the EEG data (Fig. 3C, “channel-level analysis”). Data were high-pass filtered at 1 Hz (zero-phase, 2nd order, Butterworth) to attenuate slow drift. Scalp, muscle, and noise channels were separately re-referenced using the average across channels after temporarily removing bad channels (>3 SD). To attenuate undesirable motion and muscle artifacts, we used a data cleaning algorithm called iCanClean that uses strategically placed reference channels to remove noisy EEG subspaces (Downey & Ferris, 2023). iCanClean uses canonical correlation analysis to detect and reject components from the scalp channels that are highly correlated with components from the noise and muscle channels. Further details of how iCanClean was applied to this dataset were previously described in Jacobsen & Ferris (2023). We then used CleanLine (Mullen, 2012) to remove 60 Hz line noise. We removed bad channels based on statistical criteria: probability (criterion = 5), standard deviation (criterion = 500), and kurtosis (criterion = 5) (Gwin et al., 2011). We used clean_artifacts to reject noisy time windows and channels (from clean_rawdata v2.91) (Kothe, Miyakoshi & Delorme, 2019). On average, 119 channels (±6 SD) and 95% of frames remained in each dataset. The EEG data were re-referenced again to the common average and maintained full rank. EEG, ground reaction force, and motion capture data were merged into a single dataset using a hard-wired synchronization waveform (0.5 Hz).

Next, we transitioned to source-level analysis (Fig. 3C, “source-level analysis”). We used adaptive mixture independent component analysis to decompose the preprocessed EEG data, which amounted to an average of 50 minutes/person, into maximally independent components (Palmer, Z-Delgado & Makeig, 2012). We down-sampled the component data to 256 Hz and computed an equivalent dipole model for each independent component using the custom head models and Fieldtrip software (v20210614; Oostenveld et al., 2011). After source localization, we excluded artifact components based on the following criteria: positive power spectral density slope (linear slope [2–40 Hz] >0), high residual variance (>15%), localization outside the skull, low probability (<50%) of being a brain component per ICLabel (Pion-Tonachini, Kreutz-Delgado & Makeig, 2019), and high cross-frequency power-power coupling [quantified with PowPowCat; Thammasan & Miyakoshi, 2020] in specific frequency bands that are characteristic of eye and muscle components [correlation coefficient >0.3 in low (<8 Hz) and high (>30 Hz) frequency windows].

An average of 21 (±7 SD) components per participant were then used for group clustering using k-means. We estimated the optimal cluster number using evalclusters and the average of the results from the Calinski-Harabasz, silhouette, and Davies–Bouldin methods. We used two equally weighted spatial features for clustering, scalp topography and dipole location, and excluded time-frequency features to avoid potential inflation of the false-positive rate (Kriegeskorte et al., 2009). In cases where multiple components per participant were present within a cluster, we selected the components that explained the greatest amount of variance (i.e., the lowest independent component order number) to ensure that only one component per participant per cluster was considered, thus preventing artificial inflation of the sample size. Clusters lacking a dipole for a particular participant may reflect technical limitations of mobile EEG, such as contaminated independent components that did not meet our criteria to be classified as a “brain” component.

We performed time-frequency analyses for each dipole cluster. We computed event-related spectral perturbation plots and normalized them using single trial baseline removal (average log power across time within one gait cycle), which has been shown to be less sensitive to noisy trials (std_precomp parameters: cycles = [3,0.8], pad ratio = 2, basenorm = ‘on’, baseline = NaN (manual baseline subtraction applied later), trialbase = ‘full’) (Grandchamp, Delorme & Neri, 2011). In some analyses, we applied an additional baseline removal, which are described in the subsequent sections. We linearly time-warped event-related spectral perturbations to the group median gait cycle length using foot lift-off and contact events. We computed the average of all independent components’ event-related spectral perturbations within a cluster to create grand mean event-related spectral perturbation plots. We then calculated event-related spectral power for each subcondition to compare effects of error and perturbation magnitude. We examined spectral power changes in the frequency range of 2–50 Hz, which is slightly beyond theta, alpha, and beta bands, to gain a clearer picture of the extent of any spectral fluctuations and to visually inspect the data for signs of motion and muscle contamination.

Time course of gradual adaptation

To test our hypothesis that there would be no changes in theta power across gradual adaptation, we compared event-related spectral power (4–7 Hz) between pre-gradual adaptation and subconditions of gradual adaptation. Figure 4A shows how strides were binned during gradual adaptation conditions for a sample subject, with the associated belt speeds in Fig. 4B. The ramp-up period with a constant acceleration was broken down into three stages with 30 strides each: early ramp, mid ramp, and late ramp. We refer to the first 30 strides after the belts reached a full 2:1 speed ratio as the early 2:1 split condition and the last 30 strides as the late 2:1 split condition. We averaged event-related spectral perturbation data first at the participant level and then at the group level for each cluster of interest.

We evaluated differences between pre-gradual adaptation and each subcondition (early ramp, mid ramp, late ramp, early 2:1 split, late 2:1 split) spectral power (2–50 Hz) using cluster-based permutation tests in Fieldtrip (Oostenveld et al., 2011) through EEGLAB (Delorme & Makeig, 2004), which uses the uses the Monte Carlo method to estimate the permutation p-value. Cluster-based permutation is a robust nonparametric statistic that both corrects for multiple comparisons and reduces potential for false negative effects (Maris & Oostenveld, 2007), making it a popular method for testing hypotheses in high-dimensional EEG data. For all cluster-based permutation tests, we used a significance threshold of 0.05 and 15,000 iterations. We used the maximum Cohen’s d from each significant pixel cluster as a measure of effect size (Meyer et al., 2021) and quantified the 95% confidence interval using bootstrapping (3,000 iterations). If there was an effect of condition on spectral power (2–50 Hz) in the anterior cingulate clusters, we then performed an additional cluster-based permutation test to evaluate the effect of condition (pre-gradual adaptation vs. gradual adaptation subcondition) on theta power (4–7 Hz) and test our hypothesis. As an exploratory analysis, we quantified spectral power changes (2–50 Hz) in the sensorimotor, posterior parietal, and posterior cingulate cortices and compared post-adaptation #2 subconditions (initial, early, late) with pre-gradual adaptation using cluster-based permutation tests.

Small versus large errors

We further categorized the asymmetric strides using a median split of step length asymmetry values, resulting in strides with large and small kinematic errors. On average 60 trials (range = [8,122]) of each error type went into each subject mean bin for the event-related spectral perturbation analysis. We calculated grand mean event-related spectral power across subjects during all large and small errors with respect to pre-gradual adaptation and then computed the difference in between large and small errors. We used cluster-based permutation tests to evaluate significant differences between each condition and pre-gradual adaptation and between both large and small error conditions.

Abrupt versus gradual adaptation errors

We tested if how a perturbation is introduced influences error-related electrocortical activity, when error and perturbation size are similar. We quantified differences in event-related spectral power between abrupt and gradual adaptation when the perturbation magnitude was equal and the behavioral error sizes were similar. This occurred when the belt speed ratios were the same (2:1), and step length asymmetry values were similar (within two standard deviations of the gradual adaptation asymmetric stride mean). We extracted the first 30 asymmetric strides during the full 2:1 split of gradual adaptation and matching asymmetric strides from abrupt adaptation. As an exploratory analysis, we used cluster-based permutation tests to evaluate the effect of condition (abrupt versus gradual) on event-related spectral power during these asymmetric strides. The caveat with this comparison is that the order of abrupt and gradual was not randomized because the primary focus of the study was to investigate brain dynamics during a novel abrupt split-belt perturbation (Jacobsen & Ferris, 2023). It is possible the previous practice of abrupt transition split-belt walking attenuated changes in electrocortical dynamics in addition to the effect of the gradual split-belt protocol.

Participant demographics

Prior to any processing, we discarded datasets from three participants due to issues with EEG data recording (n = 2) or protocol errors (n = 1). The group analysis included data from the remaining 30 participants (n = 30), which included 15 females and 15 males, with a mean age of 22.8 ± 2.6 years (range = [19,29]) and a body mass index of 23.70 ± 3.85 (mean ± SD). Table 1 summarizes the participant demographics.

Step length asymmetry

We assessed the effect of split-belt adaptation on step length asymmetry. One participant displayed no change in step-length asymmetry during abrupt adaptation, as determined by the slope of the linear fit. Therefore, we classified them as an outlier and removed them from in all subsequent analyses. All remaining participants (n = 29) exhibited an increase in step length asymmetry during split-belt adaptation (Fig. 2B), with smaller asymmetry occurring during gradual than abrupt adaptation (p < 0.0001). Step length asymmetry became more symmetric over time in abrupt adaptation, and more variable over time in gradual adaptation. We observed after-effects following both abrupt and gradual adaptation, as indicated by the positive step length asymmetry during the early post-adaptation periods (Fig. 2B). Gradual adaptation notably lacks the rapid change in step length asymmetry seen in the initial abrupt adaptation, post-adaptation #1, and post-adaptation #2 periods (Fig. 2B).

There was high inter- and intra-participant variability in timing of asymmetric strides during gradual adaptation. The average difference in treadmill belt speeds during the initial 30 asymmetric strides of gradual adaptation was 0.23 m/s, which means errors began to emerge before the belts reached a full 2:1 split. The belt symmetry during these initial errors was 0.55 ± 0.08 (mean ± SD).

Figure 5 shows the relationship between individuals’ normalized abrupt and gradual adaptation performance. A value of one indicates a good performance, meaning they were quick to reach steady state during abrupt adaptation or had a low percentage of asymmetric strides during gradual adaptation. The linear regression model revealed there was no correlation between abrupt adaptation performance and gradual adaptation performance (adjusted R² = 0.01, p = 0.26). Therefore, the number of strides a participant took to reach steady-state step length asymmetry during abrupt adaptation was not predictive of the percentage of asymmetric strides they would have during gradual adaptation.

Cortical clusters

The optimal k-means identified 12 cortical clusters, seven of which were included in this analysis because they were in brain regions of interest and included at least half of the participants. The location of cluster centroids within a standard brain image (Montreal Neurological Institute) are provided in Fig. 6. In Fig. 7, we provide more detailed information on each cluster, including the Talairach atlas label. Considering the inherent source localization error of EEG (∼1–2 cm according to Seeber et al. (2019); Sohrabpour et al. (2015)), we recognized that in certain cases, the Talairach atlas label may have overestimated the precision of the cluster centroid’s location within the brain. Therefore, we assigned new labels to some clusters (Fig. 7; “cluster label” column), which we will use throughout this paper. The posterior parietal cluster centroid was in the left hemisphere, but the individual sources were distributed between both hemispheres. We relabeled a cluster that the atlas called mid-cingulum as posterior cingulate. The cluster was very close to the posterior/midcingulate border, and a previous split-belt study that used a brain imaging technique with higher spatial resolution than EEG (positron emission tomography) found metabolic changes within the posterior, but not midcingulate cortex (Hinton et al., 2019).

Time course of gradual adaptation

Figure 8A displays the belt speed ratio and magnitude of group-averaged step length asymmetry during subconditions of the gradual adaptation paradigm. These data are repeated for Fig. 9A so that the step length asymmetry (kinematic errors) and electrocortical dynamics can be visualized simultaneously for multiple brain regions. Mean step length asymmetry increased from pre-gradual adaptation to the early 2:1 split of gradual adaptation (Fig. 8A). During gradual adaptation, step length asymmetry was greatest during the late 2:1 split but was still about half the size of the after-effects during initial post-adaptation #2 (Fig. 8A).

Most cortical clusters (71%) did not exhibit changes in spectral power across gradual adaptation (Figs. 8–9). Figure 8B shows grand mean event-related spectral perturbations across stages of gradual adaptation with respect to pre-gradual adaptation; red indicates increased spectral power (neural synchronization) and blue indicates decreased spectral power (neural desynchronization). All regions without significant differences from pre-gradual adaptation, as determined by cluster-based permutation tests, have a semi-transparent mask. Cluster-based permutation tests revealed that the left anterior cingulate exhibited a significant difference in gait-related spectral power (2-50 Hz) from pre-gradual adaptation during the middle of the ramp-up (mid-ramp) of gradual adaptation (p = 0.010), before the belts reached a full 2:1 speed ratio. The effect size, as measured by the maximum Cohen’s d, was d = 1.59 (95% confidence interval (CI) = [1.21,1.95]). Descriptively, this positive pixel cluster (region 1 in Fig. 8B) was located near the theta and alpha bands. There were no significant differences in event-related spectral power when step length asymmetry magnitude was greatest, during the end of the split-belt period (late 2:1 split). Among the sensorimotor and posterior parietal clusters, only the posterior parietal cortex exhibited an effect of gradual adaptation subcondition on event-related spectral power from the cluster-based permutation tests. We found a positive pixel cluster during the late 2:1 split subcondition (region 3 in Fig. 9B; p = 0.019) whose effect size was d = 1.48 (95% CI = [1.06,1.95]).

Figure 10 shows the average theta power (4–7 Hz) near the left anterior cingulate across the gait-cycle for pre-gradual adaptation and mid ramp. A cluster-based permutation test revealed that there was an effect of condition on theta power (p = 0.033, d = 0.86, 95% CI = [0.55, 1.16]). There was greater theta power during the mid ramp condition than during pre-gradual adaptation.

Visual inspection of the event-related spectral perturbations indicated that theta synchronization near the left anterior cingulate and alpha synchronization near the posterior parietal cortex increased over the course of gradual adaptation, from the early ramp to the late 2:1 split period (Figs. 8B and 9B, respectively). Theta synchronization near the left anterior cingulate cortex appeared to be strongest preceding left foot contact, during right leg single support. By the end of gradual adaptation, alpha synchronization near the posterior parietal cortex was occurring during the swing phase of the right leg in the posterior parietal cortex.

There were four cortical clusters that exhibited an effect of a post-adaptation #2 condition on event-related spectral power as determined by cluster-based permutation tests: the posterior cingulate, bilateral sensorimotor, and posterior parietal cortices. The right posterior cingulate exhibited desynchronization during the initial post-adapt. #2 (Fig. 8B, region 2; p = 0.003, d = 1.48, 95% CI = [1.03, 1.96]) and early post-adapt. #2 conditions (Fig. 8B, region 3; p = 0.023, d = 1.36, 95% CI = [0.86, 2.02]) with respect to pre-gradual adaptation. By late post-adaptation #2, the right posterior cingulate displayed synchronization (Fig. 8B, region 3; p = 0.004, d = 1.28, 95% CI = [0.87,1.90]). The abrupt transition back to tied-belt walking (initial post-adaptation #2) also evoked desynchronization with respect to pre-gradual adaptation near the left sensorimotor (Fig. 9B, region 1; p = 0.0004, d = 1.63, 95% CI = [1.18, 2.14]), right sensorimotor (Fig. 9B, region 2; p = 0.009, d = 1.40, 95% CI = [1.05, 1.82]), and posterior parietal cortices (Fig. 9B, region 4; p = 0.0028, d = 1.35, 95% CI = [0.98,1.85]) that dissipated by early post-adaptation #2. Descriptively, these enhanced desynchronizations were within the alpha and theta bands for the left sensorimotor cortex, the beta band for the right sensorimotor cortex, and the alpha and beta bands for the posterior parietal cortex.

Small versus large errors

We categorized the asymmetric strides as being a large or small error based on step length asymmetry. The magnitude of group-averaged step length asymmetry was 0.07 ± 0.02 for large errors strides and 0.04 ± 0.02 (mean ± SD) for small errors (Fig. 11A). Figure 11B displays the event-related spectral power during strides with large and small errors and the difference between the two conditions (large minus small) with respect to pre-gradual adaptation. Regions without significant differences between large and small asymmetric strides (Fig. 11B, third column) have a semi-transparent mask. There was only one significant effect found of condition on event-related spectral power, and this was near the left anterior cingulate source for strides with large asymmetry (i.e., large errors) with respect to pre-gradual adaptation (p = 0.015, d = 0.93, 95% CI [0.64, 1.70]). The location of this significant positive pixel cluster was approximately in the theta and alpha bands preceding left foot-contact (region 1 in Fig. 11B). There were no significant differences between small and large errors in any cortical cluster.

Abrupt versus gradual adaptation errors

As an exploratory analysis, we compared asymmetric strides between abrupt and gradual adaptation when the perturbation and error magnitude were approximately equal. The group-average step length asymmetry magnitude during these asymmetric strides was 0.06 ± 0.02 for abrupt adaptation and 0.07 ± 0.02 for gradual adaptation (Fig. 12A). Figure 12B shows grand mean event-related spectral perturbations from abrupt adaptation errors (with respect to pre-abrupt adaptation), gradual adaptation errors (with respect to pre-gradual adaptation), and the difference between the two (abrupt minus gradual). Cluster-based permutation tests revealed enhanced synchronization during abrupt adaptation errors with respect to pre-abrupt adaptation in the left anterior cingulate (Fig. 12B, region 1; p = 0.011, d = 1.36, 95% CI = [0.99, 1.80]) and left posterior cingulate cortices (Fig. 12B, region 2; p = 0.014, d = 1.17, 95% CI = [0.94, 1.47]). For gradual adaptation errors, only the right posterior cingulate showed changes in spectral power, which decreased with respect to pre-gradual adaptation (Fig. 12B, region 4; p = 0.024, d = 1.53, 95% CI = [1.01, 2.20]). Cluster-based permutation tests indicated that there was an effect of subcondition (abrupt versus gradual) on event-related spectral power near the left posterior cingulate (p = 0.005, d = 1.41, 95% CI [1.02, 1.91]), right posterior cingulate (p = 0.004, d = 1.45, CI = [0.92, 2.13]), and posterior parietal cortices (p = 0.017, d = 1.47, 95% CI [1.06, 1.96]). The left posterior cingulate cortex (Fig. 12B, region 3) and right posterior cingulate cortex (Fig. 12B, region 5) showed enhanced synchronization during abrupt adaptation errors compared to gradual adaptation errors, whereas the posterior parietal cortex showed stronger desynchronization (Fig. 12B, region 6).

Visual inspection revealed more alpha and beta desynchronization near the posterior parietal cortex (region 6) and more theta synchronization near the bilateral posterior cingulate cortex (regions 3 and 5) during the abrupt asymmetric strides compared to the corresponding gradual asymmetric strides (Fig. 12B). In the left posterior cingulate, there was a positive (red) pixel cluster approximately within the theta band (region 3 circled in Fig. 12B) that occurred following left foot contact in the plot showing the difference between the abrupt and gradual errors; this was the limb on the slower belt. Similarly, there was a positive (red) pixel cluster observed in the right posterior cingulate during left leg single support (Fig. 12B, region 5). While there are differences between abrupt and gradual adaptation errors in both posterior cingulate clusters, the difference in the left posterior cingulate was caused by increased theta power during abrupt adaptation, whereas the difference in the right posterior cingulate was caused by decreased theta power during gradual adaptation. Lastly, alpha desynchronization near the posterior parietal cortex appears to occur throughout the gait cycle but is more prominent during right leg stance (Fig. 12B, region 6).