Participants and Protocol
A total of 10 participants (age-matched control: N = 5 and individuals with stroke: N = 5) participated in a pilot training study using the KINARM Exoskeleton and a joystick (Fig. 1A) (28). The following inclusion criteria were used for all participants: normal or corrected-to-normal vision and at least 18 years of age. Individuals with stroke were included if they had a single, unilateral, chronic stroke (> 6 months post-stroke). The following exclusion criteria were used for all participants: previous recent history of significant upper body injury, history of a disease that may impact sensation (e.g., diabetic sensory neuropathy), and any history of a neurological disease or injury (e.g., Parkinson’s disease) (Table 1). The current study was approved by the University of Delaware Institutional Review Board and all participants provided informed consent.
Table 1
Participant demographics. Values presented next to field name indicate scoring categories (TLT) or maximum value. Abbreviations: M: Male, F: Female, R: rRght, L: Left, A: Ambidextrous, FM-UE: Fugl-Meyer Upper-Extremity Assessment, FIM: Functional Independence Measure, TLT: Thumb Localization Test, PPB: Purdue Pegboard, BIT: Behavioral Inattention Test, MoCA: Montreal Cognitive Assessment. For age, BIT, and MoCA, values are reported as median ± standard deviation. For month post-stroke, FM-UE, FIM, and PPB, values are reported as median, with minimum and maximum participant scores in brackets.
| Age-Matched Control (n = 5) | Individuals with Stroke (n = 5) |
Age – mean ± std | 64.54 ± 5.86 | 69.16 ± 8.86 |
Sex | 2 M, 3 F | 2 M, 3 F |
Dominant hand | 4 R, 1 L, 0 A | 5 R, 0 L, 0 A |
Months post-stroke | | 81 [18, 100] |
More affected side | | 4 R, 1 L |
FM-UE (maximum = 66) | | 43 [17, 65] |
FIM (maximum = 126) | | 125 [121, 126] |
TLT {0, 1, 2, 3} | | 3, 1, 1, 0 |
PPB | | 2.0 [0, 10] |
BIT (maximum = 146) | | 143.2 ± 3.03 |
Field cut | | None |
MoCA (maximum = 30) | | 26.6 ± 1.67 |
General Robotic Methods
Participants were seated in the KINARM exoskeleton with their arms at ~ 80° abduction. The lengths of the robot arm segments were adjusted to fit each participants limb length. Participants were then wheeled into the integrated augmented reality system to begin testing. For Pre- and Post-Assessments (see below), participants were seated with both arms supported by the KINARM. For the Training Protocol (see below), participants were seated with one arm in the KINARM and the opposite arm supported by a small table with the joystick secured to the table. This table was adjusted to be at the same height as the opposite arm that was supported by the KINARM (Fig. 1A).
Clinical Assessments
For participants with stroke, we used the following clinical measures to characterize upper limb function: the Upper Extremity portion of the Fugl-Meyer Assessment (FM-UE) to examine motor function of the upper limb (29), the Functional Independence Measure (FIM) to determine functional ability (30), the Thumb Localizer Test (TLT) to determine upper limb position sense status (31), the Purdue Pegboard (PPB) to evaluate upper limb and hand dexterity (32), the Behavioral Inattention Test (BIT) to screen for visuospatial neglect (33), and the Montreal Cognitive Assessment (MoCA) to screen for cognitive impairment (34). Participants with stroke also completed visual field confrontation testing to determine if visual field cuts were present (Table 1).
Robotic Pre- and Post-Assessments
To assess the effects of training, we used two KINARM robotic tasks known to quantify upper limb motor control (Visually Guided Reaching (VGR), Fig. 1B) and upper limb position sense (Arm Position Matching (APM), Fig. 1C). The methods for both tasks have been previously described in detail (8, 9, 35–37). During the VGR task, participants were instructed to make reaches to a visual target that would appear on the screen (37). Briefly, a 1 cm red target would appear on the screen at one of four locations. Participants were instructed to move their arm, where their fingertip was represented as a 1 cm white cursor, to the target and hold until the next target appeared. Participants started each trial at the center of the four locations. All participants performed VGR with both arms. During the APM task, all visual feedback of limb location was eliminated. Without vision of their arms, the robot would passively move one arm to one of four positions within the workspace and participants would actively move their arm to mirror-match the final position of their passively moved arm (8). For this task, individuals with stroke had their more-affected arm passively moved by the robot and actively matched the movement location of the robot with their less-affected arm. The limb that was passively moved was counterbalanced for age-matched controls.
Training Protocol
The objective of each training trial was for participants to use a joystick to actively guide their opposite arm to a target position. More specifically, the joystick manipulations the participants made with their less-affected arm were translated to passive movement imposed by the robot on the opposite side. Three different end target locations and four different start positions for each end target (total of 12 starting positions) were used in the robotic training session. The start positions for each target set were defined as the vertices of a square with a 20 cm edge and centered around the target location (Fig. 1D). Before the training trials started, participants were allowed four familiarization trials with the joystick and KINARM set up. Here participants were allowed to use the joystick to control the position of their hand inside the KINARM to ensure they understood the relationship between the operation of the two devices. At the beginning of each training trial, visual information about the target hand location and current hand location was provided via a cyan and white circle with a diameter of 1 cm, respectively. When the visual target appeared on the screen, the participant then used the joystick to guide passive movement of their opposite arm that was supported by the robot to the seen position of the end target. All visual information was extinguished 500 ms after the start of the trial. After the participant verbally confirmed that they felt they reached the end target, the trial ended. After each trial ended, participants were passively moved to the next (pseudorandomized) starting position by the experimenter and visual information about target and current hand location was provided to begin the next trial. The starting position sequence was pseudorandomized over a list of four starting positions for each target in order to make it extremely difficult that participants would be able to learn to perform the task by only refining their action plan that involved their less-affected arm. In that case, in fact, participants would be able to achieve task success by only recalling the exact sequence of actions to be implemented with their less-affected arm to reach the target position and without having a reliance on online proprioceptive feedback.
The experimenter-controlled joystick was implemented to 1) reduce potential proprioceptive drift in individuals with stroke, 2) ensure uniform starting positions across participants, and 3) reduce strain from passively generated movements to the motors of the KINARM exoskeleton, which would appear whenever the exoskeleton was position-controlled with gains sufficiently large to achieve smooth and accurate position control along a continuous trajectory, in the current software implementation. Participants completed 9 trials from each start position, for a total of 108 trials. There were 4 unique trajectories per target set, for a total of 12 unique trajectories overall. Overall, the training took 38 ± 5 minutes for participants with stroke and 31 ± 4 minutes for control participants.
Implementation of joystick input to control the robotic arm
The joystick signal was processed and sent to the robot computer using Simulink (Mathworks, Natick, MA, USA). The joystick signal (two 14-bit signals for x and y) was converted in a normalized range ([-1, 1]), and processed to ensure a continuous and smooth movement trajectory command signal fed to the KINARM. The joystick signal with magnitude smaller than 0.01 was filtered out using a deadband filter to eliminate noise. The filtered joystick input command was then scaled to the amount of force applied to a virtual mass-damper system (gain:1.8E-5 N). The virtual mass-damper system had a mass of 1.25 kg and damping constant of 5 Ns/m in both the x and y direction. Via a Simulink transfer function model, we calculated the velocity of the virtual mass every 1 ms, subject to the force input extracted by the processed joystick signal. The resulting velocity signals in each x- and y-direction were filtered using a moving average filter with a 10-sample window. The filtered velocity signal was then sent from the computer running Simulink to the robot computer, which directly controlled the motors of the KINARM Exoskeleton via a Simulink Real-Time model, via UDP communication in every 10 ms. To mitigate dramatic velocity command changes resulting from sampling ratio discrepancy between UDP communication (0.1 kHz) and Simulink Real-Time Model (4 kHz), a Kalman filter was implemented in the Simulink Real-Time Model. This filter estimates missing velocity commands between the velocity commands received via UDP communication.
Robotic parameters to quantify behavior
Both assessment tasks (VGR and PM) are standardized and thus have auto-generated parameters and reports. To quantify motor (VGR) and proprioceptive (APM) function, for each task we utilized a composite Task Score, as well as kinematic parameters to quantify behavior. The Task Score is a single-value composite measure of all movement parameters for the task. This value can be z-transformed to compare performance of an individual against a normative model controlling for age, sex, and handedness (38). The use of these measures has been previously described (39–45). In addition to the Task Score, the following parameters were used to quantify upper limb motor function with the VGR task: Posture Speed – median hand speed when hand is resting, Reaction Time – amount of time between end target onset and movement onset, Initial Direction Angle – angular deviation between vector from hand position at movement onset to end target and vector from hand position at movement onset and hand position after initial phase of movement (e.g., movement onset to first local minimum after max speed), Initial Distance Ratio – ratio of distance between hand position at movement onset and offset and distance covered during initial phase of movement, Speed Maxima Count – number of hand speed maxima between movement onset and offset, Min-Max Speed Difference – average difference between pairs of local speed minima and maxima, Movement Time – time between movement onset and movement offset, Path Length Ratio – ratio of straight line from hand position at movement onset and offset and actual path travelled from movement onset to offset, Max Speed – maximum hand speed between movement onset and offset, Number of No Reaction Times – number of trials when no movement onset was calculated, Number of No Initial Stabilizations – number of trials when participant did not stabilize in the start target, Number of False Starts – number of trials when movement onset occurred less than 130 ms after end target turned on, Number of No Movement Ends – number of trials when movement offset was not detected before trial end, and Number of End Targets Not Reached – number of trials when the end target was not reached (37, 46). Briefly, movement onset was calculated as the time when the cursor left the start target, and movement offset was calculated as the hand position 1.2 seconds after the peripheral target was reached (46). The following parameters are used for the APM measure: Absolute Errorx – average absolute distance in x-dimension between passive hand position and mirror-reflected active hand position, Absolute Errory – average absolute distance in y-dimension between passive hand position and mirror-reflected active hand position, Absolute Errorxy – average absolute distance in the xy plane between passive hand position and mirror-reflected active hand position, Variabilityx – average standard deviation of hand position in x-dimension for all targets, Variabilityy – average standard deviation of hand position in y-dimension for all targets, Variabilityxy – average standard deviation of hand position in the xy plane for all targets, Contraction/Expansion Ratiox – ratio of absolute difference between average x-position of left targets and right targets between active and passive arms, Contraction/Expansion Ratioy– ratio of absolute difference between average position of average y-position of left targets and right targets between active and passive arms, Contraction/Expansion Ratioxy – ratio of area moved between passive and active arm, Shiftx – average difference between mirror x-position of active arm and x-position of passive arm (positive values are lateral shift and negative values for medial shift), Shifty – average difference between mirror y-position of active arm and y-position of passive arm (positive values for distal shift and negative values for proximal shift), and Shiftxy – root-sum-squares of shift in the x-dimension and shift in the y-dimension (8, 46).
Data and Statistical Analysis
To determine if sensorimotor learning occurred the robotic training task, we first calculated and then fit trial time and end point error to a three-parameter exponential decay model. Trial time was defined as the time from the start of the trial to when the participant verbally announced they felt matched. End point error was defined as the Euclidean distance from the position when the participant verbally announced they felt matched to the desired position. We fit these data for each participant with an exponential decay model with three free parameters: Initial Error (\({E}_{0}\)), Learning Rate (\(\lambda\)), and Asymptotic Error (\({E}_{n}\)), as a function of Trial Number (t).
$$Learning value\left(t\right)={E}_{0}*{e}^{-\lambda *t}+{E}_{n}$$
This model fit was bootstrapped to improve estimation of fit parameters. To bootstrap this estimation, for 1,000,000 iterations, we re-sampled the data with replacement and fit the model. The bootstrapped estimation was then determined as the median from each bootstrapped-parameter-distribution.
To test our predictions, for each participant group, we compared Task Scores from the pre- and post-training time periods for both VGR and APM. We used directional permutation tests (\({H}_{0}:post>pre\)) with 1,000,000 permutations for these comparisons, such that we expected participants to decrease their Task Score (i.e., improve behavior) after training (47). To quantify the effect size, we used common language effect size (CLES) which describes how often a sample from one distribution will be greater than a sample from another distribution (48). We then performed this same analysis on each parameter for both tasks. For example, for the VGR task, we compared performance for Posture Speed from the pre- and post-training assessments, and for the APM task, we compared performance for Absolute Errorx during the pre- and post-training assessments. Additionally, we compared each of the three free parameters from the modified exponential decay model between individuals with stroke and age-matched controls. We used directional permutation tests for these comparisons, such that we expected individuals with stroke to show larger initial and final error as well as smaller learning rates (47). The effect size of these comparisons was also quantified with CLES (48).