This was a single-blinded, randomized controlled trial conducted at a rehabilitation hospital. Participants were randomly assigned to the active-assistive robotic intervention (using active-assistive exoskeletal robot with assistive force; ACAS) group or active robotic intervention (using active exoskeletal robot without assistive force; ACT) group in a 1:1 allocation ratio using a randomization table calculated by the NCSS-PASS program. A researcher computed the randomization sequence using the program, another researcher enrolled participants and the other researcher assigned participants to interventions. Random allocation was conducted by using consecutive sealed opaque envelopes indicating group allocation, which were placed in a plastic container in numerical order. Each group completed 20 sessions of 30-minute robotic intervention, 5 days a week, for 4 weeks, conducted by an experienced research physical therapist in a research intervention room. Additionally, both groups received 30 minutes of conventional therapy for the affected upper limb, 5 days a week, for 4 weeks. The study was approved by the institutional review boards of a hospital, and all participants provided written informed consent before enrollment. Our study was registered retrospectively with ClinicalTrials.gov (NCT03465267).
The study enrolled 20 patients with upper extremity dysfunction owing to stroke who were admitted in a rehabilitation hospital between March 2017 and December 2017. The inclusion criteria were: (1) age of >19 years; (2) presence of hemiplegia owing to ischemic or hemorrhagic stroke; (3) stroke duration of >3 months; (4) hemiplegic shoulder and elbow flexion/extension with a Medical Research Council scale score of 3 or 4 for muscle strength; (5) affected upper extremity Fugl-Meyer Assessment score (FMA) of 21–50; (6) shoulder and elbow flexor spasticity with Modified Ashworth Scale score of ≤1+; (7) cognitive function of the level that facilitates the understanding and obeying of instructions of this study; and (8) absence of a limit of range of motion of the shoulder and elbow joint, as determined by the neutral zero method. The exclusion criteria were as follows: (1) history of surgical treatment of the affected upper extremity; (2) musculoskeletal problem of the upper extremity such as fracture, contracture, and shoulder subluxation of more than two finger breadth; and (3) cybersickness, that is, occurrence of nausea or vomiting while seeing a screen.
Active-assistive robotic intervention group
In the ACAS group, we administered intervention using an Armeo® Power (Hocoma Inc, Zurich, Switzerland) (Figure 1A), which is a three-dimensional exoskeletal active-assistive robot used for upper extremity rehabilitation. Actuators actively assist affected arm movement as established extent, on top of arm weight support offsetting the device weight. Participants were trained with game-based virtual reality environment with focus on proximal upper limb movement.
Active robot intervention group
In the ACT group, we used an Armeo® Spring robot, (Hocoma Inc, Zurich, Switzerland) (Figure 1B), which is an exoskeletal active robot for three-dimensional upper extremity rehabilitation. The Armeo® Spring provides gravity compensation, offsetting the device and participant’s upper extremity with the help of a spring but not with robotic actuators. Participants were trained under the same virtual reality environment as were those included in the ACAS group.
We evaluated FMA to measure impairment, Wolf Motor Function Test (WMFT) to measure activity, Stroke Impact Scale (SIS) to measure participation, according to the International Classification of Functioning, Disability, and Health (ICF) concept.  To determine more detailed kinematic outcomes, smoothness and mean speed were measured. Outcome measures were checked at baseline (T0), after 2 (T1) and 4 weeks of the intervention (T2), and 4 weeks after the end of the intervention (T3).
The primary outcome measure was WMFT, which quantifies the upper extremity functional activity using 15 functional tasks. WMFT-score is rated on a 6-point scale, with the score ranging from zero to five; thus, the total score ranges from 0–75. WMFT-time is the sum of the time required to perform all 15 tasks. The higher the WMFT-score or the shorter the WMFT-time, the better the motor activity.
Secondary outcome measures were FMA score, SIS score, and kinematic data. FMA score, which ranges from 0 to 100, is a quantitative indicator of motor impairment following stroke, with a higher scores reflecting a lower impairment.  We used FMA-UE (shoulder, elbow, forearm, wrist, and hand; 33 items, 0-66) and FMA-prox (shoulder, elbow, and forearm; 18 items, 0–36). SIS version 3.0, which is a stroke-specific, self-reported questionnaire, has been applied as a health-related quality of life measurement tool to assess participation. [13, 14] We measured eight domains of SIS (strength, hand function, ADLs and instrumental ADLs (ADLs/IADLs), mobility, communication, emotion, memory and thinking, and social participation); the score of each domain ranges from 0 to 100; the higher the score the better the health status. In the present study, four domains (strength, physical, ADLs/IADLs, and social participation) that are more relevant to proximal upper extremity function were selected for secondary outcome assessment. We also determined SIS-overall (sum of scores of all eight domains) and SIS-function (sum of scores of ADLs/IADLs and social participation).
With regard to kinematic analysis for detailed information about impairment, we recorded the position of affected upper extremity using the trakSTARTM system (Ascension Technology Corp, USA), which measures the movement of an electromagnetic sensor tracing 6 degrees of freedom (x, y, and z axes) at 80 Hz of sampling rate during each reaching movement. In the present study, the sensor was attached at the distal phalanx of the middle finger with double-sided tape, and the wire was fixed to the skin with bandage; the reference transmitter was located behind the participant (Figure 2). Each patient was asked to sit in a chair in front of a table, the height of which was adjusted such that the elbow is flexed at an approximate angle of 90° in the sagittal plane; however, the distance of the table from the participant was maintained such that comfortable reaching is ensured. Participants practiced the reaching task three times to be familiarized with the experimental setup, which is described as follows. Buttons (base button and three target buttons) were positioned according to each participant’s affected arm length (from the distal end of the middle finger to the acromion). Three target buttons were set on a vertical wooden plate in front of the participant at the height of the participant’s xyphoid process and at a distance of 75% of the arm length in three different positions on the transverse plane (ipsilateral, central, and contralateral). The central button was installed in front of the midline, and two other buttons (ipsilateral and contralateral button) were placed in the ipsilateral and contralateral position at an angle of 45° from the central button. Base button was placed on the table in front of the midline at 25% of the measured arm length. Subsequently, participants were asked to reach from the base button to one of the three different target buttons, subsequently bringing back the upper limb to the base button at their own comfortable speed. Those movements were repeated nine times (three times to reach each target button in a randomized order) with 1 min of rest between each movement. Patients were instructed to limit trunk movements without a trunk restraint.
Subsequently, two kinematic performance indices were computed on the basis of the position data during reaching: spectral arc length (SAL) and mean speed (MSP). SAL is a dimensionless measure reflecting the smoothness, which was calculated using the arc length of the Fourier magnitude spectrum of a movement speed profile.  A higher SAL value indicates a smoother and thus a better movement. It is also known as an important marker reflecting motor recovery of patients with stroke.  MSP was calculated by dividing the distance of actual trajectory by the time required for reaching from the base button to each target button.
We assessed the usability of the patients with stroke with on the basis of individual interview at the end of the intervention. Usability was also determined on the basis of interviews conducted by the research physical therapists, who were in charge of the robotic intervention, and physiatrists, who observed the robotic rehabilitation at the end of the present study.
We analyzed the participants who completed outcome measurements at T2 at the least. When the results of T3 were not measured, the last observation carried forward method was used; thus, missed outcomes at T3 were filled in with those determined at T2. For the comparison of baseline characteristics between two groups, Fisher’s exact test and Mann-Whitney U test were applied for categorical variables and continuous variables, respectively. Repeated measures of analysis of variance (RM-ANOVA) were conducted using the group (ACAS or ACT) and time (T0, T1, or T2) to compare the effect of each intervention across time, and time × group interactions were assessed. Greenhouse-Geisser corrections were applied when the violation of sphericity occurred. Additionally, Mann-Whitney U test was performed for the intergroup comparison of kinematic data. A p-value of <0.05 was considered statistically significant. All statistical analyses were performed using IBM SPSS Statistics for Windows, Version 20.0. Armonk, NY: IBM Corp.