Subjects
This was a two-center RCT conducted in Hong Kong between 2017 and 2019. Sub-acute stroke survivors were screened and recruited from two local hospitals: Hospital S and Hospital T. This study was approved by the Institutional Review Board of the hospitals and was designed following the principle of the Declaration of Helsinki. All recruited subjects read and signed consent form before participation.
Recruited subjects satisfied the following inclusion criteria, including (1) first episode of stroke within 2 months, (2) hemiparesis resulting from unilateral ischemic or hemorrhagic stroke, (3) ability to walk with one-person assistance (functional ambulatory category, FAC ≥ 1), and (4) sufficient cognition to follow instructions and understand the content and purpose of the study. Subjects were excluded if he/she had (1) uncontrolled cardiovascular or respiratory disorders, (2) moderate to severe contractures in lower extremities (modified Ashworth scale, MAS > 2 at ankle, knee, or hip), or (3) orthopedic or muscle disorders that affected mobility.
Intervention
Recruited subjects were randomly allocated into three groups by drawing lots: (1) power-assisted ankle robot (PAAR), (2) swing-controlled ankle robot (SCAR), and (3) conventional training (CT). All subjects received conventional rehabilitation protocol (physiotherapy and occupational therapy) prescribed by rehabilitation team of the inpatient training centers for 2 h/weekday, including standard lower-limb exercises on standing, balance, stepping, and walking.
For subjects who were assigned in PAAR and SCAR, 30-min robot-assisted training (at least two sessions/week, total 20 sessions) were integrated into their conventional training routine (2 h/weekday) without time compensation. Each robot-assisted training session consisted of 10-min over-ground walking, followed by 10-min stair training (ascending/descending), then another 10-min over-ground walking. The two training centers had similar settings: having staircase with handrail (5–10 steps with 120–150 mm step height, 1.2–1.5 m width, 350–400 mm depth) and long corridor (≥ 10 m) cleared of obstacle with minimal turning. Subjects were free to take break anytime but resting was also counted in the training time. The whole session lasted around 45 min including robot setup (don/doff) time. A trainer walked beside the affected side of the subject and held subject’s waist belt all the time to ensure safety. The trainer administered verbal cue on head/trunk extension in case of increased trunk kyphosis, or mid-line awareness when subjects leaned on the unaffected side. Subjects used their own walking aids prescribed by the hospital rehabilitation team, including walking cane, quadruped stick, and walker. The rehabilitation team checked the subject’s vital sign and reviewed his/her functional capability before each session. The trainer regularly asked the subjects if they felt any pain and discomfort during training. The number of stairs and walking distance covered were documented in each session as a record of training intensity and capacity.
Exoskeleton ankle robot
Both PAAR and SCAR were provided with the same exoskeleton ankle robot but in different operation mode adjustable by the trainer. The robot was worn inside subject’s footwear on the affected side throughout each robot-assisted training session (Fig. 1). The wearable robot was modified from an articulated AFO with the ankle joint coupled with a rotatory servomotor (Dynamixel MX-106R, ROBOTIS, South Korea) and a torque amplifier (1:1.67 gear ratio) that can provide powered assistance in ankle dorsiflexion/plantarflexion directions. The robot can identify changes in foot loading and gait phases using embedded force sensitive resistors (FSR-402, Interlink Electronics, USA) placed under heel and forefoot. An inertial measurement unit (MPU6050, 6-axis MotionTracking, InvenSense, USA) mounted on the shank can measure leg tilting angle for classifying user walking intention on level and stair walking [27]. The robot weighted 0.5 kg (including AFO and motor) on the leg, with the control box (0.5 kg) held by the trainer.
The ankle robot in PAAR mode was intended to provide powered ankle assistance together with residual motor function to facilitate over-ground walking and stair training. If the robot detected walking intention in either over-ground walking or stair ascending, the servomotor generated sufficient constant torque on the affected ankle in dorsiflexion direction to prevent foot drop and to facilitate foot clearance with around 10° ankle dorsiflexion throughout swing phase of walking, until heel strike was detected and then ankle joint was free to move in stance phase. Contrary, if the subject was stair descending, the servomotor generated constant torque in plantarflexion direction to facilitate loading response when the affected foot was landing on the lower step, then the ankle joint was free to move when the heel touched the floor. To calibrate dorsiflexion assistance, subjects were told to perform voluntary maximum ankle dorsiflexion on the dropped foot, while the motor torque gradually increased in dorsiflexion direction until the paretic ankle reached 10° dorsiflexion. To calibrate plantarflexion assistance, subjects were told to stand quietly on both leg while the motor torque increased gradually in plantarflexion direction until the torque was sufficient to uplift the heel to 10° plantarflexion on affected side. The calibration was performed by the trainer at the beginning of each session to adjust for any progression of functional changes throughout the 20-session gait training. The calibrated ankle torque requirement matched with previous research 3.6 ± 2.4 Nm on stroke subjects (n = 80) with mild spasticity (MAS ≤ 2) [18].
The ankle robot in SCAR mode acted as a swing-controlled orthosis, which switched between locked and unlocked ankle joint based on the gait phases [14]. Whenever the robot detected terminal stance as the foot was lifted up from the ground, the ankle joint was locked by the servomotor in the neutral position to prevent foot drop condition during swing phase for foot clearance [24], effectively acted as a rigid AFO. When heel strike and foot contact with the ground were detected, the servomotor released the ankle joint to allow unimpeded forward ankle rocker during stance phase. Similar passive swing-controlled AFO had been proposed by previous researches showing these devices were able to prevent foot drop and enhance gait stability [14].
Outcome measures
Clinical assessments were carried out by blinded assessors within a week before the intervention (Pre) and within a week after the intervention (Post). The same assessor administered both Pre and Post assessment of a subject. All assessors were blinded to group allocation. Clinical assessments were selected based on a meta-analysis that aimed to evaluate the effectiveness of wearing AFO, which recommended outcome measures targeting on mobility, walking, and balance [16, 29]. All clinical scores were assessed on subjects without using any assistive devices, neither the ankle robot nor any orthosis subjects wore.
The primary outcome measure was FAC, which was used to classify gait independency based on a six-point scale, ranging from FAC = 0 “needs help from at least two persons to walk” to FAC = 5 “can walk independently anywhere, including uneven surfaces and stairs”. Previous study determined that FAC ≥ 4 could predict community ambulation at 6-month with 100% sensitivity and 78% specificity after 4-week rehabilitation [4, 13].
The secondary outcome measures included Berg balance scale (BBS) and timed 10-m walk test (10MWT). BBS was used to assess static and dynamic balance ability based on 14 functional tasks with varying difficulty, including sitting, standing, transfer, reaching, stepping, and turning. Each task was rated on a five-point scale, ranging from 0 to 4 based on the performance of the subject in completing the activity. The highest BBS score was 56, while the score of 45 had been shown to be a cut-off score for greater functional independency and lower fall risk for stroke survivors [29]. 10MWT measured the self-selected walking speed in meter per second over a short distance. The uses of walking aids and manual assistance were documented and made consistent for Pre and Post assessments. Studies indicated walking speed had good correlation with functional independency and disability level [13, 30]. Stroke subjects who walked with self-selected speed > 0.4 m/s were considered at least limited community ambulators [31].
Statistical analysis
The power analysis for sample size calculation was based on our previous RCT that investigated the effects of robot-assisted gait training on functional independency of chronic stroke survivors [28], with the between-group difference in FAC score had effect size 0.471. The estimated sample size for the current study was 48 for three groups with 0.8 power (1-β) [32]. The power analysis was performed using G*Power version 3.1.9.6.
The statistical analysis aimed to evaluate any significant difference between robot-assisted trainings and conventional training on sub-acute stroke survivors. All outcome measures were analyzed based on the intention-to-treat principle, which used the last-observation-carried-forward method to impute the last available data to missing entries for any drop-out. Analysis of covariance (ANCOVA) was used to compare the improvement (Post) scores in FAC, BBS and 10MWT between groups. To reduce the expected confounding effect of the variation in baseline clinical scores, we adjusted the group means using baseline (Pre) scores as covariate. If ANCOVA revealed significant effects, post-hoc comparison between groups were tested using Mann–Whitney U-test for ordinal variables (FAC and BBS) and independent samples t-test for continuous scales (10MWT). To explore the practical significance of group differences, effect sizes were calculated as follows:
$$Effect Size = {{\left( {Mean_{Group1} - Mean_{Group2} } \right)} \mathord{\left/ {\vphantom {{\left( {Mean_{Group1} - Mean_{Group2} } \right)} {SD_{Pooled} }}} \right. \kern-\nulldelimiterspace} {SD_{Pooled} }}$$
The established criteria of the effect size, which reflects the treatment effect within the target population, were small (< 0.41), medium (0.41 to 0.70), or large (> 0.70) [33]. Statistical results were reported with the effect size in 95% confidence interval (95%CI). Two-tailed level of significance set at 5%. Statistical analysis was performed using IBM SPSS Statistics Version 23 (IBM Corp., USA).