Novel gait training alters functional brain connectivity during walking in chronic stroke patients: a randomized controlled pilot trial

A rotational treadmill was designed to provide turning-based treadmill training [2]. This treadmill is similar to a regular treadmill except for its circular running motor belt (0.8-m radius), which forces patients to continually turn rather than walk straight. Participants walked on the perimeter of the circular belt as it rotated either clockwise or counterclockwise. Participants were trained in both directions, with the affected leg as the inner limb for 15 min and then as the outer limb for 15 min with a 5-min break in between. The order of walking with the affected limb as the inner or outer leg was alternated in each consecutive session. All participants wore an unweighted safety harness to prevent falls. Participants were also allowed to place their hands on the handrail for balance support but were instructed to refrain from holding the handrail. As progressively faster speeds are needed to continue challenging the locomotor abilities of individuals with hemiparesis [10], the treadmill speed, which began at each individual’s comfortable turning speed on level ground, was increased by increments of 0.05 m/s every 5 min as tolerated. After completing the turning-based treadmill training, a 10-min ambulation training followed [2].

The control group received training on a standard treadmill (Biodex, Shirley, NY). Other than the type of treadmill, the training protocol was the same as that described for the experimental group. Training speed was initially set at the individual’s comfortable walking speed on level ground and increased by increments of 0.05 m/s every 5 min as tolerated. Participants were trained in two 15-min phases, followed by a 10-min ambulation training as in the experimental group.

The motor system consists of well-tuned interactions of excitatory and inhibitory mechanisms in the cortical and subcortical areas. Many researchers have used coherence analysis, a method of power spectral cross correlations between spatially separate neural systems, to reflect the functional communication between two systems, i.e., coupling between two brain regions (EEG-EEG coherence) [6], or between cortical commands and consequent muscle activation (EEG-EMG coherence) [7]. However, the conventional coherence methods show linear dependency and may be insufficient to study the nonlinear signals in biological and physical systems, such as data pertaining EEG and EMG [11], especially if the signals are contaminated by noise or the oscillatory frequency band is not carefully defined [12, 13].

In the present study, the time-frequency cross mutual information (TFCMI) method was used to solve the problems mentioned above, to calculate the mutual information between two temporal power sequences within a task-specific band between two neural systems [14, 15], and thus to examine the treatment-induced brain reorganization during walking. To obtain consistent information, in both the experimental and control groups, subjects walked on the regular treadmill for 5 min while the neurophysiological measures and the functional connectivity of the signals from different scalp locations (EEG-EEG connectivity) and the brain and muscle (EEG-EMG connectivity) were assessed .

EEG, EMG, and footswitch signals during treadmill walking were simultaneously recorded from Ag/AgCl electrodes using a 40-channel QuickAmp amplifier (32 EEG channels, 4 bipolar channels for EMG and 4 auxiliary channels for footswitch) and Brain Vision Recorder software (Brain Products, Gilching, Germany).

Using the international 10–20 system, 32 EEG electrodes located over the frontal-parietal areas were used (F5, F3, F1, Fz, F2, F4, F6, FC5, FC3, FC1, FCZ, FC2, FC4, FC6, C5, C3, C1, CZ, C2, C4, C6, CP5, CP3, CP1, CPZ, CP2, CP4, CP6, P1, Pz, P2, POz). Electrodes were recorded against an average reference calculated by the amplifier hardware. Horizontal and vertical electrooculography (HEOG and VEOG) signals were also recorded. EEG and EOG signals were amplified with a bandpass of 0.16–50 Hz, recorded in DC mode and sampled at 1000 Hz. Electrode impedances were kept below 5 kΩ [16].

For recording EMG, the skin was shaved and cleaned with alcohol swabs to reduce impedance before applying 11 mm Ag-AgCl surface electrodes (Brain Products, Gilching, Germany). Surface electrodes were placed on the motor point of the tibialis anterior (TA) of the affected leg, which is located at 1/3 on the line between the tip of the fibular and the medial malleolus [17]. Electrode impedances were kept below 5 kΩ.

Four footswitches were placed respectively under the heel and the big toe of each leg and taped to the sole of the foot using medical tape to measure the “heel strike” and “toe-off” events of the gait cycle. Signals from footswitches were recorded using the QuickAmp amplifier and Brain Vision Recorder software to synchronize signals from footswitches with those of the EEG and EMG. Significant coupling between EEG and EMG signals using the conventional EEG-EMG coherence method was found approximately 24–40 Hz during the swing phase of walking in healthy human subjects [8]. Therefore, the target gait event in this study was focused on the swing phase. The swing phase was defined as the time interval between toe-off and the subsequent heel strike of the same foot.

EEG-EEG connectivity and EEG-EMG connectivity were calculated using the TFCMI method. The data recorded within the target gait event was off-line processed as follows. The independent component analysis (ICA) algorithm was first applied to decompose EEG recordings into 15 independent components. Among these independent components, the signals that were highly correlated to the VEOG or HEOG (with correlation coefficients higher than 0.6) were rejected from the reconstruction to eliminate the potential contamination from eye movement during recording. The EEG/EMG recordings were then segmented into epochs measuring from 200 sample points before the onset of the “toe-off” to 300 sample points after the onset of the “toe-off”. The epochs with a signal fluctuation more than a standard deviation from the mean signal value were then discarded to remove potential motion artifact and ensure signal stability. The stable EEG signal over time within the same patient was obtained through ICA-based artifact rejection. The Butterworth zero-phase filters with pass band from 1 to 50 Hz were applied. Each window of filtered EEG/EMG data across channels was transformed into the time–frequency domain using the Morlet wavelet transformation to obtain temporal spectral (power) map [18]. The corresponding Morlet wavelet transformation was given by time–frequency maps encompassing the alpha (8–12 Hz) and gamma (24–40 Hz) activities that were created separately within each window. The power across the selected frequency bands in each channel was averaged to produce power curves of the target bands, i.e., alpha and gamma bands in this study. The temporal series of the averaged power signals of the target bands in each channel were used to compute the cross mutual information between any two channels in each window. A full mathematical description of TFCMI has been published previously [14]. The TFCMI value, which is an index of functional connectivity that estimates the relationship between two channels on their averaged power changes over the target band, was calculated based on the entropy and joint entropy. The TFCMI between the i^th and j^th channels were calculated as follows:

H: the entropy; F: the averaged power signals at the channel; P: the probability density function (pdf); b: the index of sampling bins used to construct the approximated pdf.

The cortical maps of EEG-EMG connectivity were presented in individual channels (32 channels). Considering pairwise connectivity, a 32-channel network for the EEG-EEG connectivity may produce (32 × 32–32)/2 possible functional connections, which may be convoluted when visualized. Therefore, the 32 channels were grouped into 13 regions to better visualize the EEG-EEG connection between cortical regions (Table 1).

The gait speed and symmetry measure independent features. Gait speed provides the overall gait performance and may differentiate levels of disability, while the temporal gait symmetry indicates differences of single support time between two limbs due to the difficulty of bearing weight on the affected leg during the stance phase or in advancing the affected leg during the swing phase [19, 20]. The speed and temporal asymmetry ratio at a comfortable walking speed were obtained from the GAITRite system (CIR system, Inc., Havertown, PS); these two gait parameters were selected based on the results of our previous study [2]. The concurrent validity [21] and reliability [22] of the GAITRite system for stroke subjects have been established. The GAITRite walkway was 4.75 m long and 0.9 m wide, and the pressure-sensitive area was 4.30 m long and 0.61 m wide. The contact time and location of each footfall were recorded and the gait parameters were calculated on a laptop with application software. Participants began their trials 1.5 m before the mat and continued walking for 1.5 m beyond the end of the mat to eliminate the effect of acceleration and deceleration [22]. Participants were allowed to use their customary ankle-foot orthosis or assistive devices [22]. The assistive device and/or ankle foot orthosis used at the preintervention assessment were also used at the postintervention assessment. Fragmented steps and unrelated marks from assistive devices were removed automatically by the GAITRite software or manually by the investigators [22]. The average of three trials was used for data analysis [22]. The temporal asymmetry ratio was calculated using the following formulas [2]: