Robotically driven Error Augmentation training enhances post‐stroke arm motor recovery

Stroke is a major cause of long‐term functional disability and requires physical rehabilitation. Due to population aging, the number of people post stroke is going to rise. Robotically neurorehabilitation has a great potential to improve outcome measurements. Error Augmentation training using a robotic interface is thought to promote motor recovery by enhancing proprioceptive feedback, which motivates and challenges patients to optimize their performance during training. Here, we investigated the effectiveness of robotic Error Augmentation training on motor recovery after a stroke, compared to standard robotic training in a null field. Post‐stroke patients were randomly assigned to one of two groups: a study group (n = 9) that was trained on a 3D robotic system applying Error Augmentation forces, and a control group (n = 7) that carried out the same protocol in null field conditions. The robotic rehabilitation intervention was applied in addition to the standard rehabilitation protocol of the rehabilitation center. Error Augmentation training increased clinical scores compared to standard robotic training by 266% on the Motor Assessment Scale, and 88% on the Fugl‐Meyer scale. The Motor Assessment Scale scores were significantly correlated with the Fugl‐Meyer scores (p = 0.03, r = 0.541). There were more movement errors on the initial trials of the game sequence using the DeXtreme robotic device with Error Augmentation compared to trials with no force field. This difference vanished however after 10 trials. Error Augmentation training decreased the number of movement units and jerkiness compared to the control treatment. There was a robust effect of magnifying the acceleration component of movement using EA on the smoothness of the movement. These findings suggest that EA training may enhance motor performance possibly through motor adaptation. Future study should include EMG to better elucidate the neural mechanisms involved in motor learning post CNS injury.


INTRODUCTION
Rehabilitation after a stroke using interactive robotic technologies provides automatic, repetitive training, with high volume and accuracy that can be harnessed to objectively measure patients' motor performance. 1Robotic devices can also be used in rehabilitation to deliberately interfere with patients' movements by providing mechanical perturbation or by distorting visual feedback during movement execution. 2,3While both of these methods involve a large amount of practice, their physiological rationales differ considerably.5][6] Sensory dysfunction may also impact motor capacities, as a result of abnormal processing of proprioceptive feedback, cutaneous and pressure information from the upper limbs, 7,8 and tactile loss, which may also affect coordination and dexterity. 9Functional deficits stemming from these motor and sensory impairments are often compensated for by over-activating muscle groups that are not recruited in normal circumstances. 10Using compensation strategies to accomplish a daily task is appropriate from a functional standpoint but does not necessarily enhance genuine recovery from motor impairment. 11espite the surge in virtual reality and robotic technologies in the last two decades, a large proportion of stroke patients still suffer from significant long-term disabilities. 12Robot-mediated rehabilitation has the potential to improve outcome measurements of stroke survivors. 13Two main forms of robotic therapy have been developed to enhance recovery after a stroke.In robotic-assisted therapy, the device helps patients complete the motor task when they are unable to do so by themselves.This can involve supporting the arm during movement or guiding the movement toward the target using external forces. 14This type of robotic training mimics daily activities while minimizing the use of compensation strategies.However, this type of therapy may not maximally motivate the patients to make a sufficient effort to promote motor plasticity.][17][18] Error Augmentation (EA) training was developed to overcome these drawbacks.During normal adaptation, the proprioceptive, cutaneous, and visual feedback perceived during movement execution are utilized to calibrate the motor output on the next trial.In cases where sensory afferent inputs are affected, such as after a stroke, augmenting these inputs has been hypothesized to enhance patients' error correcting capacity.In EA training, the robotic device exerts perturbation forces on the hand that pushes it away from a straight trajectory line, and thus augments the movement error.The rationale is based on the idea that provoking movement errors will provide more appropriate sensory inputs and hence contribute to adaptation. 19n the current study, post-stroke patients were randomly assigned to two intervention groups that were engaged in robotic therapy.The control group underwent robotic therapy in a null field environment, and the study group underwent a similar protocol with EA.The findings show that in the EA group, increased movement deviations in the first 5-10 trials vanished by the last trials of the game.When applied in addition to standard clinical rehabilitation, EA therapy significantly increased scores on both the Motor Assessment Scale (MAS) and Fugl-Meyer (FM) as compared to the group trained in a null field environment.Further, the improvement in these clinical scores were significantly correlated.
the study.Exclusion criteria were complete hemiplegia, Parkinson's disease, multiple sclerosis, Alzheimer's disease, or other neurological diseases, sensory aphasia or upper limb apraxia and significant visual or hearing deficits, significant shoulder pain, or inability to provide informed consent.

Apparatus
For this study, we used an upgraded version of the apparatus described in a previous paper 20 and the device along with the study design is resented in Figure 2. Briefly, BioXtreme Ltd. is an Israeli private company (Company number 514435866) has developed a robotic arm coupled with adaptive algorithms for the treatment of upper limb post-stroke rehabilitation, utilizing a 3D VR environment to motivate patients.The device is FDA and CE registered, and its methodology and technology are globally patent protected.The apparatus was composed of a robotic arm with three segments: one rotating horizontal plate attached to the base of the apparatus, a second plate, which was attached to the horizontal base plate and measured 500 mm in length, and the third segment which was attached to the second one on one side and to a gimbal handle at its extremity, measuring 400 mm in length.The robotic arm weighed 45 kg.The robotic motors were programmed to eliminate the weight of the robotic arm in the null field environment or have specific EA forces.Three sensors tracked the hand position every 20 ms.Hand-reaching movements were carried out in a three-dimensional workspace.

Procedure
This study was conducted at the Bait-Balev Rehabilitation Center, Nesher, Israel, was approved by the local Institutional Review Board, and was registered at Clinicaltrials.gov# NCT03578250.Patients who had a stroke were screened for eligibility for the trial and provided their signed informed consent.Block randomization of four participants used a web-based computer program.The investigator who administered the robotic treatments was only informed about the allocation of each participant immediately prior to the first treatment.Each participant was evaluated on two clinical assessment  scales, the Fugl-Meyer scale (FM) and the Motor Assessment Scale (MAS), at two time points: once before the first intervention session (T1) and again after the last session (T2).Clinical assessments were carried out by an investigator who was blind to the group allocation of the participants.Participants were blind to their group allocation throughout the study.Patients were allocated either to the control null field (n = 7) or the EA study (n = 9) group, with an allocation ratio of 1:1. Figure 2A depicts the intervention protocol.Participants in both groups were given 2 weeks of training consisting of three 20-min sessions per week.Two different game applications were used as an intervention protocol in both groups: the Market Stand app game and the Alchemist app game (manufactured by BioXtreme Ltd., Petach Tikva, Israel).The two initial treatment sessions only included one game application, the Market Stand, given the complexity of the other game, the Alchemist.The other treatment sessions included both Alchemist and the Market Stand game play.Each session was composed of about 6 games in total (for example, four Market Stand games and two Alchemist games).The first game in each session, in both apps., was conducted without a force field, whereas the other games were performed either with or without the force field, depending on group allocation.Each game consisted of 10-15 trials.

Hand reaching task in three-dimensional space
During the experimental sessions, the participants were comfortably seated in a chair and held the robotic handle while viewing a visual screen (Figure 2B).In the Market Stand app., an avatar hand represents the hand of the participant, and a bee is the target.By moving the robotic arm, the participants see the movement of the avatar hand on the screen.On each trial, a bee flies to the starting point at the bottom of the screen and stops at this point.Participants bring their arm to the starting point where the bee is located.When the participants' hand reaches the starting point, the bee moves to a random point on the screen and stops.The screen then shows a straight line between the avatar's hand and the bee, which represents the optimal trajectory to reach the bee.The goal is to swat the bee.Attempts were considered successful if the participants executed the hand reaching movement that placed the avatar's hand on the bee within 4 s.Misses were defined as failing to complete the trajectory within this time frame, or if the hand deviated from the straight trajectory line by more than 45 • during the initial third of the movement.Each time the avatar hand swatted the bee, a flare went up and the straight line disappeared.Then, another bee moved to the starting point for the next trial.The inter-trial interval was 4 s.On-screen visual feedback indicated the runtime of the game, the number of successful trials, the weighted movement-error scores and the number of successful attempts.
The Alchemist game involves pouring a vial full of liquid into another container without spilling it.At the start of each trial, the participants need to move an empty vial to the starting point, which is then filled with boiling liquid while it is held above a flame.The participants hold the cup above the flame until the vial is completely full, which takes about 3 s, and then wait for the Go signal.The participants then need to move the vial in position to pour the liquid in the vial into the collecting vessel that could be located at different distances and directions, within a time limit of 5 seconds.Trials are said to be successful if there is minimal spillage of the liquid during the movement and the liquid is poured into the container.

Force field algorithm
Two different force field algorithms were applied for the two game applications.The first algorithm was applied on the Market Stand app.and was described in detail in a previous paper. 21Briefly, EA forces were applied perpendicular to the straight trajectory line and at a distance from it.The magnitude of forces was dependent on the magnitude of the error, that is, the distance of the hand from the straight trajectory line.Additionally, forces were dependent on the distance of the hand from its starting point.For the first factor, forces were increased as the error magnitude increased.For the second factor, forces were decreased as the hand location moved away from the starting point, so that the effect of the increased lever as the hand moved away from the body was minimized.
In the Alchemist app., the force field algorithm was applied to increase the error of the acceleration component aligned with the direction of the hand movement.Specifically, participants needed to avoid jerkiness of their hand movements when moving the vial away from the flame toward the collection vessel.In other words, changes in movement acceleration during the reaching movement affected the smoothness of the translation movement.In the case of excessive acceleration, the algorithm pushed the hand forward (away from the body) to elicit a pulling reaction by the participant.When the hand slowed down in the case of excessive deceleration, the algorithm pulls the hand toward the body to elicit a pushing reaction from the participant.

Quantification of movement errors and jerkiness
Movement errors on the Market Stand app.were expressed as the deviation of the hand from the straight trajectory line connecting the starting point and the hand endpoint of movement at 500 ms from movement onset.Quantifying the movement errors at 300 or 400 ms led to similar results.Movement onset was defined as the first point at which the hand had reached 10% of its maximal velocity.The movement deviations between groups were compared for baseline levels and for the experimental trials.To obtain the baselines, deviations from the first games in all sessions were compared between groups (Figure 4D).For the experimental trials, (minus the first game in each session) force field versus null field performance was compared (Figure 4E).Two measures were used to quantify the smoothness of the hand reaching path on the Alchemist app.The Number of Movement Units (NMU) was computed as the number of zero-crossings in the acceleration trace, which corresponded to the number of peaks in the velocity trace. 22,23The Normalized Jerk Score (NJS) was used to overcome the divergence of the jerk score as the movement duration increased and was calculated as follows: where r ′′′ is the third derivative of hand position, that is jerk, t is the movement duration and l is the movement amplitude. 24Accordingly, the NJS quantifies the jerkiness of the hand during movement, normalized to the movement amplitude (range of movement) and duration (time).

Clinical outcome measures
As mentioned earlier, the main objective of the study was to evaluate the clinical effect of EA treatment, on the motor capacities (mostly impairment more than disability).Reliable indication on functional and clinical improvement, may not solely rely on kinematic measures output from the robotic interface.Therefore, we decided to test the results of the study by using also two familiar, and well documented clinical tests, assuming that using it may help us to establish the trial results.The Fugl-Meyer 25 is designated to classify the impairment status of patients post stroke, based on predefined recovery stages.It is widely used by clinicians, and validated mainly through construct validity studies that correlate the score with those of other clinical scales.This upper extremity motor function scale consists of 33 items scored on a 3-point scale, where 0 corresponds to an inability to complete the test item, 1 represents partial ability, and 2 represents full completion, resulting in a maximal total score of 66 points.The test items assess reflexes, the capacity to move in and out of synergy, the ability to isolate movement to the shoulder, elbow, and wrist, and to grasp various objects.Higher scores indicated less limb impairment; lower scores indicated more limb impairment. 25he Motor Assessment Scale, upper extremity 26 This test is composed of three parts that assess upper arm function, hand movements, and advanced hand activities.Each contains six functional tasks that aim to assess similar skills, but with increasing difficulty.The difficulty can involve executing the task with/without gravity, statically versus dynamically, and dexterity in different ranges of motion.The maximal score on the MAS is 18 points.The MAS is highly reliable for assessing patient post stroke, both in the acute and chronic phase, 27 and both between different assessors and with the same assessor in different occasions.It was also validated when compared to other scales such as FM. 28

Statistical analysis
Each patient was assessed on the FM scale and the MAS before the first intervention (T1) and after the last intervention (T2).The changes in clinical scores between groups were first calculated separately for patients with a high level of performance (high skill) and patients with a low level of performance (low skill), based on their clinical scores, and then between the two complete groups.On the FM scale, the cutoff score was 41 and for the MAS the cutoff score was 8 points.
For each measure and each group and at each skill level, the mean scores at T1 were subtracted from the mean scores at T2 and divided by the mean scores at T1 (to obtain the percentage improvement) (Figure 3A,B).Next, we quantified improvement in clinical scores in the force field (FF) group, using the control group scores as a reference.For each clinical scale and each skill level, the mean score of the FF group was divided by the mean score of the control group.For example, in the MAS, for the low skill subgroups, the mean score of the control group was 61% and the mean score of the FF group was 112%; hence the FF group improved by 112/61 = 182% relative to the control group.After completing this step for the two subgroups with the two clinical scales, we applied a weighted average of these results to evaluate the relative improvement for the complete groups, without differentiating between skill levels.For example, if in the low skill subgroup, the FF group with two patients improved by 182% relative to the control group, and in the high skill subgroup the FF group with seven patients improved by 418%, the improvement for the whole group was calculated as (182 × 2 + 418 × 7)/9 = 366% relative to the control group (Figure 3C, MAS).
Differences between clinical scales were assessed by Pearson correlations (Figure 3D).Effects were considered significant at p < 0.05.To assess the effect of EA training on hand deviations (Figure 4D,E) and kinematic measures (Figure 5) a two-way ANOVA was calculated with group and time as the main effects.

Error Augmentation training improves motor performance as compared to standard robotic training
To compare the groups on the clinical scales, the participants were divided into high and low skill levels.The cutoff score for the FM scale was 41 points and 8 points for the MAS. Figure 3A,B depicts the comparisons between all subgroups for the two clinical scales.On the FM scale, patients with low skill levels improved by 37.4% in the control group and 32.1% in the FF group.In the high skill subgroup, the control patients improved by 4.6% compared to 11.1% in the FF group.On the MAS, patients with low skill levels improved by 61.5% in the control group and 112.5% in the FF group.In the high-skill subgroup, control patients improved by 6.8% compared to 28.4% in the FF group.Figure 3C illustrates the improvements in clinical scores of the FF group, using the control group scores as a reference.On both clinical scales, EA training resulted in a greater effect on post-stroke motor performance than standard robotic training.EA training led to an improvement of 188.4% on the FM scale and 366.2% on the MAS compared to the control group.The clinical scores were correlated (p < 0.05, r = 0.541) (Figure 3D).

Error Augmentation force field induces adaptation
Figure 4A-C illustrates hand movements in 3D, and velocity as a function of time.Figure 4D,E depicts the differences in hand deviations from the straight trajectory line between groups.Figure 4D shows the similarity in baseline performance between groups (two-way ANOVA, p > 0.05).However, exposing the FF group to an EA force field (Figure 4E, blue trace) deflected the arm from the straight trajectory line, causing increased hand deviations compared to the control group (red trace).Repeated exposure to the same force field led to a gradual decrease in arm deviations until deviations became comparable.A two-way ANOVA revealed a significant main effect for group (p = 0.0378), but not for trial number (p > 0.05).Differences between groups were significant at the beginning of the training game (p = 0.047) but not at the end of the game (p > 0.05), indicating the significant convergence of hand trajectories toward the straight line in the study group.
Figure 5 illustrates the effect of EA training applied to amplify the error of the acceleration component of movement.In this paradigm, when the movement of the hand was overly accelerated, the algorithm pushed the hand away from the body to trigger an opposing response aiming to adjust the smoothness and stability of the movement.
Figure 5A-D provides an example of two trials, by the same participant from the same game (patient number 10 from the study group).On the early trials (black trace), movements were less smooth, as indicated by the wavy pattern of the hand position (Figure 5A), the increased number of peaks in the velocity profile (Figure 5B), increased number of zero crossings in the acceleration profile (Figure 5C), and the increased number of minima in the jerk profile (Figure 5D).Later trials, on the other hand (gray trace) evidenced opposite movement kinematic properties.Summing the results for the two groups revealed a significant main effect for group (p < 0.05) and movement number (p < 0.0001) on the number of movement units (Figure 5E), with a significant interaction effect between group and movement number (p < 0.0001).EA training had no significant main effect for group on NJS but a significant main effect on movement number (p < 0.0001) and a significant interaction effect between group and movement number (p < 0.0001).EA training had a significant main effect for group (p < 0.0001) and movement number (p < 0.0001) on movement time and a significant main effect between group and trial number (p < 0.0001).Taken together, Figure 5 illustrates the robust effect of magnifying the acceleration component of movement using EA on the smoothness of the movement.

DISCUSSION
This study investigated whether EA training would result in better motor performance after a stroke than standard robotic therapy.Our results revealed significant improvements in clinical scores with EA training, compared to standard robotic training with no force field.Training with the EA force field initially deflected the arm from the straight trajectory line, causing larger hand deviations than training in the null field.At the end of the games, however, the hand deviations for the study group were similar to the deviations in the control group.Similar results were also obtained when using the EA algorithm on the acceleration component of movement, where EA decreased movement jerkiness, as indicated by the NMU and the NJS.Our findings thus support the use of EA to improve arm motor performance during the acute to subacute phase after a stroke, mainly but not limited to patients with moderate motor impairments of the arm.
0][31] Two studies reported better clinical effects, 29,30 and another reported better adaptation effects. 31Other studies have compared adaptation in healthy participants who were trained either in a control environment or with EA and found a greater learning effect with EA. [32][33][34] In post-stroke individuals, augmenting proprioceptive feedback has been shown to improve movement smoothness and elbow extension. 35However, it remains unclear whether EA training can result in greater gains than other robotic training methods since the reported clinical effects on improving motor recovery post-stroke have mainly been moderate. 29,30his study highlights the positive effect of augmenting movement errors to enhance motor recovery over standard robotic training.Nevertheless, the underlying mechanism related to these effects calls for further elucidation.Learning models suggest that the extent of movement correction depends on the magnitude of the error experienced in the previous movement. 36,37Recently, it was suggested that sensitivity to the error experienced in the previous movement increases as the magnitude of the error decreases. 38In other words, small movement errors are more reliable for correcting the ongoing motor command than large errors.Similarly, since the error sensitivity increases with small errors, there should be rapid adaptation with small but not large errors.
After a stroke, however, sensory deficits in proprioception and cutaneous afferents may modify the patients' sensitivity to errors that would be expected in healthy participants.In this case, stroke patients' threshold to errors should be higher than in non-stroke individuals, such that large errors should also result in high sensitivity that allows for the correction of movement errors.In addition, it is likely that non-neural mechanisms, such as motivation and maximal exertion during training, also contribute to positive clinical effects in EA training.
This study emphasized clinical improvements over scientific insights, while aiming to provide the shortest treatment session that resulted in recovery.Accordingly, each game consisted of about 10-15 trials, which curtailed our ability to statistically evaluate the adaptation process.The small sample size made it difficult to dissociate patients with different levels of sensory impairments, which is crucial when attempting to determine whether different levels of sensory deficits affect error sensitivity and the extent of adaptation.Finally, although the adaptation effect was found to be generalized to neighboring targets in space, 39,40 in the current study, the wide range of movement directions and amplitudes in three-dimensional space, along with the small number of repetitions, made it difficult to track adaptation within a given game.Addressing these methodical issues in future studies could shed light on the underlying mechanisms of EA that enhance motor recovery after a stroke.

Future work
Our findings only represent individuals who had suffered moderate stroke; thus, they cannot be generalized to individuals who had sustained severe upper limb disability or to individuals demonstrating other neurological diseases such a Parkinson disease and multiple sclerosis.Future study should include electromyography to better elucidate the neural mechanisms involved in motor learning post CNS injury.

CONCLUSION
Our findings confirmed that robotic rehabilitation using EA training was effective in improving functional reach movements in terms of functional tasks efficiency, smoothness, and trajectory.We think that our findings support the implementation of EA training among post stroke patients.

CONFLICT OF INTEREST STATEMENT
Authors have no conflict of interest relevant to this article.

DATA AVAILABILITY STATEMENT
Data openly available in a public repository that does not issue DOIs.

F I G U R E 2
Study design (A) showing the session's protocol including the game type used in each session, and the intervention type within a session.(B) The DeXtreme robotic device, as positioned facing the game screen.

F I G U R E 3
Comparison of clinical scores between groups and correlations between clinical scales.(A, B) Comparisons of clinical scores between groups split into high and low-skill patients.(C) Comparisons of clinical scores for the full groups.(D) Illustrates the correlation between improvement in MAS and FM.

F I G U R E 4
Hand movements and deviations from the straight trajectory line on the Market Stand app.(A) Movements were executed in 3-dimensional space.The velocity profile (B) and placement as a function of time (C) are depicted.(D) Comparisons assessed the differences on the first game in all sessions, where movements were performed in the null field in both groups.(E) Comparisons were applied in games where groups were either exposed to a null field (control) or a force field (study).

F I G U R E 5
Comparisons of kinematic measures on the Alchemist game between groups.(A-D) An example of two movements by the same patient and the same game (pt10_D3_G3) to illustrate the differences in kinematic parameters between a trial that was executed at the start of a game (black traces) and a trial late in the game (gray traces).The early trial with an increased number of peaks in the velocity profile (B), more zero crossings in the acceleration profiles (C), and minima in the jerkiness profile (D), as indicated by the Normalized Jerk Score (NJS) (D).(E-G) Comparisons of the mean number of movement units, NJS, and movement time between the two groups.