Introduction

Stroke is the leading cause of long-term adult disability and reduced ankle dorsiflexor (DF) strength, or foot drop, affects 2030% of stroke survivors [1]. The ankle joint is an integral link between the limb and environment and loss of ankle strength after a stroke results in decreased walking endurance, temporal asymmetry and reduced gait velocity [1-3]. Hemiparetic gait compensations for DF weakness are characterized by poor foot clearance and impaired midstance stability and contribute to an increased metabolic cost of walking and fall risk [2,4,5]. Management is limited to passive external support via an ankle foot orthosis (AFO) and active gait training, if employed, is laborintensive and does not deliver the timing, assistance, and intensity necessary for motor learning [4,6,7]. These methods may improve safety and speed but do not result in a sustained therapeutic effect characterized by functional independence when not worn [6,8-10].

Emerging adaptive assist-as-needed impedance control [11], modular wearable exoskeletal robot devices, and integrated sensor systems [12-14] can automate ankle training during gait and provide high intensity repetitive locomotor practice with somatosensory input. In this manner, experience-driven motor learning and neuroplasticity [15] can be maximized. Recommendations on the most effective control strategy and robotic ankle rehabilitation program is unclear, however, due to the heterogeneity of robot devices, small sample sizes and limited randomized controlled trials [16]. An unanswered question is whether the integration of robot-assisted neuromotor ankle control with treadmill training can positively impact the multi-faceted task of overground (OG) walking and have carryover when removed. We conducted a randomized controlled trial to investigate the benefit of precisely timed and graded robotic DF assist within the context of treadmill training to promote human-robotic cooperative locomotor learning across a broad population of individuals with stroke deficits. We hypothesized that treadmill-integrated ankle robot (TMR) training would improve our primary outcomes of DF and walking speed more than treadmill training alone (TM) and would have durable benefits while not wearing the robot 6-weeks and 3 months after training completion. This paper reports the comparative effectiveness of TMR versus TM based on unassisted (non-robotic) gait outcomes of peak paretic ankle dorsiflexion (DF) swing angle, DF angle at foot strike and self-selected OG gait velocity.

Materials and Methods

This was a parallel group randomized controlled trial utilizing a single blind where assessors were blinded to group assignment. Recruitment and informed consent procedures followed approved practices by the University of Maryland, Baltimore Institutional Review Board and the Baltimore Veterans Affairs Research and Development Committee and the study was conducted in compliance with all ethical practices and guidelines. Randomization using permuted blocks in two strata occurred after baseline testing defined by baseline gait speed where speeds ≥0.5 m/s separated fast walkers from slow walkers. The study statistician sent the concealed computer-generated allocation to the study coordinator upon each assignment via e-mail using study identification number. Subjects were expected to participate 3-times a week in their randomized 6-week gait training program.

Subjects

Recruitment occurred between September 2015 through April 2019. Forty-five stroke survivors (28 males and 17 females) met all eligibility criteria and were randomized after baseline data collection to either treadmill robot training using the ankle robot (TMR) or treadmill training alone (TM). Inclusion criteria was as follows: (1) index stroke > 2-months prior to enrollment with residual lower extremity hemiparesis (2) indications of hemiparetic gait and symptoms of foot-drop assessed by clinical observation of poor foot clearance during swing phase and/or gait compensations of increased hip and knee flexion, lower extremity circumduction, or vaulting; (3) not participating in physical therapy; and (4) ability to walk on a treadmill with handrail support. Individuals with unstable angina, heart failure within the last 3 months, poorly controlled hypertension, a recent hospitalization for a severe medical condition, orthopedic or chronic pain, a history of orthopedic related gait problems or severe aphasia limiting informed consent were excluded from the study. All participants signed informed consent and underwent medical evaluations to establish eligibility (Figure 1).

Figure 1: Consort diagram.

 Data collection

Assessments were performed in the research lab over a two-day period at baseline, after six weeks of training, and at two retention time points six weeks and three months after training completion by trained research staff blinded to subject randomization and not involved in the intervention. Day one included three Timed 10-Meter Walk Tests (10MWT) over an instrumented gait mat (GAITRite, CIR Systems, Havertown, Pa) to calculate spatiotemporal outcomes of mean gait speed (cm/s), stride length (cm), cadence (steps/min), and relative paretic single support and double support (%-cycle) times. Use of an assistive device (single or multi-point cane) was allowed and the average of the three walks determined self-selected OG walking velocity. Day two included Vicon supported threedimensional kinematic gait evaluations of the primary paretic ankle DF angle outcomes. The three-dimensional kinematic calculations relied on retro-reflective markers on the anterior and posterior iliac spine, lateral mid-thigh, lateral mid-gastrocnemius, lateral aspect of the foot, the great toe and heel of each leg. Neutral stance alignment or “zero” angle was confirmed based on the lumbosacral (L5/ S1) joint, bilateral anterior superior iliac spine, knee joint, ankle joint, and feet before all walking trials. All kinematic variables were expressed with respect to this neutral stance or “zero” angle. Once captured, participants walked with and without the robot across a 7.3-meter-long walkway at the baseline visit. A one-time baseline robot walking assessment calculated robot-wearing OG walking velocity to guide initial treadmill speed parameters for the TMR participants. Additional seated unassisted robot-based ankle metrics and positional data were collected for all time points as described elsewhere [17]. To minimize fatigue, participants had a two-day rest between the walking assessment days.

Intervention

All participants were supervised throughout the one-hour training with rest breaks as needed for five walking trials to achieve 30-40 minutes of activity per lab training session. The TM group’s initial treadmill speed was matched with the baseline 10MWT and the TMR group’s initial treadmill speed was matched with their baseline robot-wearing OG walking speed. Participants were encouraged to increase their treadmill speed or duration at each session and the targeted work intensity range was between 13-to15 (“somewhat hard” -to- “hard”) on the Borg Rating of Perceived Exertion Scale [18]. Training intensity was advanced within this guideline over 18 sessions (3x/week; 6 weeks) and stayed within prescribed heart rate and blood pressure thresholds set at the prestudy training cardiac stress test. AFO’s were allowed as needed for the TM training group and removed for robot application in the TMR robot-assisted group.

For the TMR training, a 3-degree of freedom (DOF) wearable ankle exoskeleton (Anklebot; Interactive Motion Technologies; Watertown, MA) with 2-DOF actuation (DF-plantarflexion, inversion-eversion) assisted ankle DF during the treadmill walking as described in the literature [19]. This ankle robot, weighing less than 3.6 kg, had two key fundamental attributes: back-drivability, a feature of the actuators to allow the robot to “get out of the way” of the user based on user performance; and impedance control for gentle human-device assist-as-needed interaction. In brief, the robot commanded DF angles and assistance to normalize foot DF with assist-as-needed re-adjustments in the gait cycle based on a performance-based progression (Figure 2). The robot parameters were re-set and individualized at every session using pre-training ankle range of motion (ROM) and spatial-temporal gait cycle values from a 30 second unassisted robot treadmill walking warmup trial. Robotic DF swing angle was guided by the warm-up trial ROM and set between 5°-9°. Paretic leg swing and stance cycle time was manually calculated through observation (e.g., average time over 10 strides for the same event) during this warm-up trial. Initial swing and heel-off-to-toe-off percentages were set in this manner; if not available nominal values of 35-40% and 20-25% were assigned for swing and stance respectively. Robotic assistance for the gait sub-events were precisely timed using insole microswitch sensors (Myopac Jr., Run Technologies, Mission Viego, CA). The robot dynamically modulated the DF robotic output (e.g. assist levels) for “human-informed” robotic actuation in the training session [19,20]. The training treadmills did not offer body weight support but were equipped with a support harness for safety in the event of loss of balance. The robot set-up included an adjustable shoulder strap worn by the user to offset the robot’s weight and provide anti-gravity support through the swing phase of walking. A minimum of six sessions defined training participation based on the motor learning profile of the unassisted peak paretic swing ankle by Forrester et al. where at least 6 sessions were required for subjects to attain 80% of their steady-state post-training unassisted peak paretic swing ankle value [21].

Figure 2: Treadmill integrated ankle robot training.

Sample size

The target sample size was 72 (36 per group). The sample size was calculated based on a two-sided 0.05-level two sample t-test where 36 participants per group provided 85% power to detect a difference between groups if the mean dorsiflexion changes differed by 0.72 standard deviations (i.e., an “effect size” of 0.72). However only 59 were assessed for eligibility in our ~3.5-year study period due to enrollment difficulties related to preexisting medical conditions, limitations in transportation, and need for family support/assistance.

Data analysis

The analysis was pre-specified, and we report here on gait performance outcomes indexed by peak paretic DF swing angle, DF angle at foot strike, and gait velocity. The groups were compared with respect to the gait outcomes using a longitudinal regression model with outcomes measured at four time points (baseline, after six-weeks of training, and the two retention time points of six-weeks post-training completion (RT1), and three-months post-training completion (RT2). The model allowed for different variances at each time point in each group, and an unstructured within-person correlation pattern [22] and was fit by restricted maximum likelihood. An advantage of this approach over repeated measures ANOVA is that it makes fewer assumptions about the variance structure, it allows for inclusion of those with missing data at some time points, and it implicitly imputes missing values. Due to the randomization, the model incorporated the assumption that the groups were equivalent in expectation at baseline.

The primary analysis was based on the principle of intention to treat (ITT). However, since some of the participants did not contribute any data (baseline or follow-up) for the primary outcomes, these participants were not included in the primary analysis, making this a modified ITT analysis. In a secondary analysis, we compared groups defined by the treatment they received. In this “as treated” (AT) analysis, we excluded those who did not participate in at least six exercise sessions and we crossed one patient over who had been randomized to receive TM, but actually received TMR. In an exploratory analysis, we restricted the analysis to those who satisfied a biomechanical foot drop definition where peak DF swing angle was less than 0⁰ at the swing phase of the gait cycle. Eight subjects (6 randomized to TMR; 2 randomized to TM) met this definition (see supplement).

Results

Fifty-nine subjects were screened for study enrollment and forty-five proceeded to randomization: 22 to TMR and 23 to TM. The randomized groups were not significantly different in all baseline categories however, more TMR participants used AFOs (61% versus 36%) and 35% were more than six years post-stroke compared to 18% in TM group (Table 1). Seven subjects (5 TMR and 2 TM) did not have baseline or follow-up primary outcome measures and did not allow for imputation therefore these subjects were excluded from the primary (modified) ITT analysis. Among those included in the modified ITT analysis, one subject was randomized to the TM group but received the TMR intervention. This individual was switched to the TMR group for our subsequent “as-treated” analysis. In addition, four (1 in the TMR group and 3 in the TM group) did not participate in at least six exercise sessions and were later excluded from the as-treated efficacy analysis. Of note, all those who participated in at least six sessions received ten or more sessions. Reasons for study attrition included loss of interest (3), transportation issues (1), aggravation of pre-existing knee pain (2), and new onset cardiac issues (1).

Table 1: Participant characteristics.

Primary outcome results are presented in Table 2. Due to randomization, this modified ITT model assumed equivalence in expectation at baseline for all groups. Based on this analysis, no significant week-6 post-training kinematic ankle DF or gait velocity differences were found. The mean post-training peak paretic DF swing angle and DF angle at foot strike was (4.84 p=0.32; 4.24; p=0.73) and (-0.70 p=0.14; -0.46 p=0.14) in TMR and TM respectively and this was neither statistically significant nor did it reach a clinically significant 5-degree increase [4]. Retention ankle DF gains were also not statistically significant. The 6-week mean post-training gait velocity was 0.54 m/s (p=0.0030) and 0.56 m/s (p=0.0008) in TMR and TM respectively, representing similar within group gain at week-6 without between group significance (p=0.48). The within group velocity improvement was significant and durable through retention (3-months post-training completion) in both groups.

For the as-treated cohort, no significant between group difference in ankle DF kinematics and gait velocity was seen. TMR had a within group trend for ankle DF angle gains at foot strike (p=0.07), but no values reached statistical significance. Significant within group gait velocity increases occurred in TMR and TM and were sustained at the three-month retention visit. In a post-hoc analysis of a small subset of subjects meeting strict biomechanically defined foot drop criteria (n=8; 6 TMR, 2 TM), therapeutic improvement in functional ankle DF was seen within group in TMR participants after 6 weeks of training but the small sample size precludes detailed statistical analyses or conclusions (see supplemental data).

Table 2: Model-based estimated means¹ of the three primary measures by treatment for the primary analysis.