Participants and Initial Screening
A total of 20 people over a 3-months period after a unilateral stroke were included in this study; the characteristics of these subjects are shown in Table 1. Participants had to be able to stand and walk independently or under supervision (Functional Ambulation Categories, range, 3 to 5). Based on a clinical assessment, we excluded individuals with a history of other neurological disorders (except stroke) and musculoskeletal disorders that affected walking capacity, efficiency, and endurance. Written informed consent was obtained from all participants before entering the study. The study procedures were approved by the ethics committee of the Samsung Medical Center Institutional Review Board (Approval Number: 2016-07-093).
Experimental Design and Equipment
This study was designed as a crossover randomized controlled trial. All participants completed a familiarization session by walking on a treadmill and their preferred walking speed was recorded. None of the participants had difficulty walking on the treadmill during this initial walk. All participants were further acclimated to the GEMS-H through a single training session of 30 min with a licensed physical therapist. For fNIRS recording, participants were assigned to two consecutive tasks: a treadmill walking task at a self-selected speed (a) with the assistance of the GEMS-H (GEMS-H) or (b) without GEMS-H (NoGEMS-H) assistance. The experiment began with a fixed standing condition (60 s), followed by one of the two walking conditions (60 s) and then a resting condition (60 s) for five repetitions (block design) (see Figure 2A). All participants were given specific instructions not to talk or laugh during testing and the participants rested by sitting for 10 min between the two tasks .
An fNIRS imaging system (NIRSscout® system, NIRx Medical Technology, Berlin, Germany) with two different wavelengths of 760 and 850 nm was used to record changes in oxygenated hemoglobin (oxyHb) concentration. The fNIRS optodes consisted of 16 LED light sources and 16 detectors, and a total of 49 useful source-detector channels were used for monitoring the hemodynamics of the bilateral primary sensorimotor cortex (SMC), premotor cortices (PMC), supplemental motor areas (SMA), and prefrontal cortices (PFC). The cranial vertex (Cz) located beneath the 1st source was the marker for ensuring replicable placement of the optodes. After the Cz position was determined on the participant’s head, an fNIRS head cap was placed on the participant’s head. The fNIRS head cap was designed to be compatible with the International 10-20 system and the interoptode distance was 3.0 cm. The fNIRS data were continuously acquired at a sample rate of 3.91 Hz through NIRStar Software version 14.2 (NIRx Medical Technologies LLC, Berlin, Germany) in MATLAB (The Mathworks, USA), which allowed oxyHb signals to be visualized in real time during data collection.
Data Preprocessing and Analysis
Changes in oxyHb concentration during two different tasks (GEMS-H and NoGEMS-H) were analyzed by the nirsLAB® software version 2017.06 (NIRx Medical Technologies LLC, Berlin, Germany) in MATLAB. Cortical regions assessed included SMC (Brodmann area 1, 2, 3, and 4, medial), PMC (Brodmann area 6, lateral), SMA (Brodmann area 6, medial), and PFC (Brodmann area 9). We used oxyHb concentration as a marker for cortical activation because oxyHb is more sensitive indicator of brain activity during human locomotion-related activities than deoxygenated hemoglobin (deoxyHb), and there was a task-related increase of oxyHb concentration in the SMC without significant changes in deoxyHb concentration [19, 22]. OxyHb has been shown to have a higher signal to noise ratio associated with scattering of light through the scalp, skull, and inactive brain tissue . For easy comparison, brains of patients are left-right flipped in the data preprocessing stage so that the stroke lesion of each subject were localized to the right hemisphere.
The fNIRS data were preprocessed to delete experimentally irrelevant time intervals from data, to remove motion artifacts, and to apply bandpass frequency filter to exclude experimentally irrelevant frequency bands. Using the components of Data Preprocessing available in nirsLAB®, discontinuities and spike artifacts of signals obtained from the 49 channels were removed and replaced by the nearest signals. The fNIRS signals were bandpass-filtered (low-cutoff frequency 0.01 Hz and high-cutoff frequency 0.2 Hz) to eliminate the effects of heartbeat, breathing, and low-frequency signal drifts for each wavelength . The acquired fNIRS signal can contain various noises that can be classified as experimental errors, instrument noise, and physiological noise. The experimental errors and instrumental noise are not related to the brain activities, so they were eliminated prior to converting the raw optical density signals into a change in oxyHb concentration, and the preprocessed signals were then converted to relative concentration changes in oxyHb using the modified Beer-Lambert law for each source-detector channel [23, 24]. Finally, the oxyHb concentration changes were averaged over 5 repetitions for each walking condition to improve the signal-to-noise ratio .
From the processed fNIRS signals, oxyHb concentration was averaged per region of interest (ROI) (i.e., bilateral SMC, PMC, SMA, and PFC) . The SMC was assessed with the medial parts of the posterior channels (channels 1, 9, and 42 in the left hemisphere and 3, 27, and 32 in the right), the SMA was assessed with the medial parts of the middle channels (channels 11 and 12 in the left hemisphere and channels 24 and 25 in the right), the PMC was assessed with the lateral parts of the middle channels (channels 10, 13, 45 in the left hemisphere and channels 23, 26, 36 in the right) and the PFC was partially assessed with channel 16 in the left hemisphere and channel 20 in the right hemisphere (see Figure 2B) [19, 22, 27]. In this study, to analyze cortical activation, task periods were divided into an early and late phase. The period between 1 and 30 s of the task was defined as the early phase to reflect the immediate hemodynamic response for walking. The period between 31 and 60 s of the task was defined as the late phase to reflect continuous brain activity during walking as Lu et al.  described in the previous study. The initial and final 5 s of each task period were excluded to eliminate the transient periods between hemodynamic responses . Block designs with a task period of 20–30 s are commonly used for fNIRS studies [29-31], but in this study, a longer task period (60 s) was used to investigate cortical activation. For quantification of activation between the serial measurements in two different tasks, we calculated ΔoxyHb in each channel, defined as oxyHb during Task Period – oxyHb during Rest Period.
Wearable Hip-assist Robot, GEMS-H
The GEMS-H was developed at the Samsung Advanced Institute of Technology (Samsung Electronics Co., Ltd., Korea) as a wearable Hip-assist Robot with an assist-as-needed algorithm for stroke patients with gait disorder. This robot was designed to deliver active-assistance torque to the both hip or hip joint of the paretic side for extension and flexion. The GEMS-H has a lightweight (2.8 kg), comfortable and slim design that can be adjusted to fit the user’s body (Figure 1). For more information regarding the strategy of assistive algorithms used for the GEMS-H, please see our previous paper .
All statistical analyses were performed with SPSS version 22.0 (IBM, Armonk, NY, USA), and the significance level was set at 0.05. Descriptive statistics are expressed as mean ± standard deviation (SD) of the mean. Brain activation during each walking condition and phase was identified as a significant increase in oxyHb concentration by performing independent t-tests with false discovery rate (FDR) correction of multiple comparison for 49 channels. Within each walking condition, paired t-tests were used to compare activation in ipsi- and contralesional hemispheres.