Mapping upper-limb motor performance after stroke - a novel method with utility for individualized motor training

Background Chronic upper limb motor impairment is a common outcome of stroke. Therapeutic training can reduce motor impairment. Recently, a growing interest in evaluating motor training provided by robotic assistive devices has emerged. Robot-assisted therapy is attractive because it provides a means of increasing practice intensity without increasing the workload of physical therapists. However, movements practised through robotic assistive devices are commonly pre-defined and fixed across individuals. More optimal training may result from individualizing the selection of the trained movements based on the individual’s impairment profile. This requires quantitative assessment of the degree of the motor impairment prior to training, in relevant movement tasks. However, standard clinical measures for profiling motor impairment after stroke are often subjective and lack precision. We have developed a novel robot-mediated method for systematic and fine-grained mapping (or profiling) of individual performance across a wide range of planar arm reaching movements. Here we describe and demonstrate this mapping method and its utilization for individualized training. We also present a novel principle for the individualized selection of training movements based on the performance maps. Methods and Results To demonstrate the utility of our method we present examples of 2D performance maps produced from the kinetic and kinematics data of two individuals with stroke-related upper limb hemiparesis. The maps outline distinct regions of high motor impairment. The procedure of map-based selection of training movements and the change in motor performance following training is demonstrated for one participant. Conclusions The performance mapping method is feasible to produce (online or offline). The 2D maps are easy to interpret and to be utilized for selecting individual performance-based training. Different performance maps can be easily compared within and between individuals, which potentially has diagnostic utility. Electronic supplementary material The online version of this article (10.1186/s12984-017-0335-x) contains supplementary material, which is available to authorized users.

The participants leaned their forehead on a headrest and viewed a horizontal mirror, approximately 10 cm below eye level ( Figure S1). A 32 inch computer screen was mounted horizontally 24 cm above the horizontal mirror so that the participant could see the virtual image of the display congruent with the plane of motion of the vBot handle. The mirror concealed the handle and the participant's hand from direct view, but the handle location was accurately indicated on the display by a cursor (a small red disc 0.3 cm radius). The participant's body midline was aligned with the centre of the display. The height of the entire workstation could be adjusted individually.
The two degree-of-freedom robotic manipulandum permits planar movements of its end-point handle, with minimal back-drive friction and inertia, across a workspace of approximately 80x45cm [1]. A foot-operated safety switch was continuously depressed by the experimenter to activate the robot when the participant was ready for each session; release of the footswitch immediately cancelled any forces generated by the motors. The vBot handle position, velocity and all applied forces (motor torques, limited to maximum of 100N) were recorded at high precision (sampling rate 1000 Hz). At the beginning of each trial a start position was indicated by a small white disk (0.3 cm) and the vBot gently guided the participant's hand towards it using a minimumjerk-based impedance controller [2]. Once the participant reached the start position the robot maintained the hand there with a spring-like resistance until a target was

Robot assistance algorithm:
We adopted a revised version of the assistance algorithm that was developed by Krebs et al. [3]. At any moment during the attempted movement the robot could provide forces along two orthogonal axes: 1. Assist forces were provided along the start-to-target axis. The magnitude and sign of Assist force at any moment depended on (i) the difference between the actual and expected progression along the time-constrained minimal-jerk-based trajectory [2] and (ii) whether or not the attempt to move was compromised by high muscle tone.
The allotted movement duration was set individually during an initial tuning session (see below). At any moment the actual hand location was compared to the expected location (based on the minimum jerk trajectory defined by that duration). At any moment, if progression towards the target was delayed (compared to expected), a damped spring-like Assist force was applied towards the target. The stiffness constant depended on whether or not the initial delay exceeded a criterion indicating high muscle tone ( 12 and 6 Ncm -1 , respectively). High muscle tone could lead to very strong resistance to the intended movement ( Figures 3B and 3C, left, main article). On the other hand high tone could also lead to very fast, rebound-like movement typically from an extended elbow posture when the robot, holding the hand at the start position, released the hold to initiate the trial ( Figures 3B   and 3C, right, and Figure 4, main article). If the initial movement towards the target was abnormally fast then a negative Assist force acted to slow the movement and promote its expected smooth progression (stiffness constant: 12 Ncm -1 ). No Assist force was provided for movements which were faster than allotted time but still within the range of normal voluntary movements (i.e. which would lead to 5cm displacement within 200ms or longer; estimated using minimum jerk). Additionally, target overshoots were also restrained by the robot. The robot controllers were triggered upon the initiation of the movement, or 2 seconds after target onset in rare occasions when the participant failed to start. In all cases, to minimize the risk of transient harmful movements a weak damping was used during the intra -trialinterval, and the introduction and termination of assistive and guiding forces were ramped up over a period of 100 ms.

Robot-assistance parameter tuning session
The allotted movement time (t a ) and the Guide force's stiffness constant (k sw ) were set individually during an initial tuning session, adopting Krebs et al's staircase method [3].
Parameters were adjusted after each cycle of 32 trials. Movements compromised by high muscle tone were excluded from the adjustment algorithm. Briefly, the rule for adjustment of t a in cycle n+1 was: ( ) ( ) ( ) and the rule for adjustment of k sw was: Usually, tuning of the assisted parameters approached a plateau within a single session (10 cycles, or 320 trials). In rare cases there was a need of an extra session. Following the tuning session, the robot-assisted parameters (t a and k sw ) were fixed for the particular individual. All subsequent robot-assisted sessions (performance mapping and/or training) then used these fixed parameters.

PM2 and PM3 parameters
The definition of PM2 and PM3 parameters was adopted from [3]. Here we provide only a brief description of these parameters, since the focus of this paper is on performance mapping procedure rather than on a specific performance measure. The PM2 and PM3 parameters measure the ability to move and to aim, relative to individually-set performance Performance mapping sessions were used for creating PM maps that then allowed selection of movements in the following training sessions (see below).
Note that since both PM2 and PM3 parameters are defined relative to individually-set parameters, they are not directly comparable across subjects. Note also that PM3 is a purely kinematic measure and therefore may be biased due to the fact that the movement is assisted; in other words the kinematics are affected by the robot guidance and assistance Assist and Guide vs PM2 and PM3 maps: The Assist and Guide parameters are useful portrayals of impairment, with minimal performance-assistance confounding issues (see main article and PM3 to select trained movements.

Impairment-based proportion of movement selection based on PM2 and PM3 maps
Individuals differ in their ability to aim and to complete movements. Hence, the selection of training movements should reflect the performance map that best captures their individual impairment. Accordingly, the number of training conditions selected from each map (PM2 or PM3) was weighted in proportion to the mean performance over the worst 25% of each map, such that more training conditions were selected from the map that showed worse motor performance. Specifically, for each map PM i (i=2,3), the mean performance ̅̅̅̅̅̅̅̅̅ was computed across the lowest 25% of the scores of the map. The ̅̅̅̅̅̅̅̅̅ score was then ranked as mild (W PMi =0), moderate (W PMi =1) or severe (W PMi =2) (see 1 ). Movements to form a training set were then chosen at random from two sets of 102 movements selected based on the steepest gradients of PM2 and PM3, to satisfy the numerical weighting: ,