Multivariate outcomes in a three week bimanual self-telerehabilitation with error augmentation post-stroke

Abstract

We present the outcomes of a study on stroke patients in a 3-week intervention of bimanual self-telerehabilitation. This training is similar to an upper-extremity treadmill in that patients can make use of their healthy arm to provide a cue for the more impaired arm. We further inspected a cohort that covertly received error augmentation training while they practiced. Finally, we focused here on the many quantitative measures available from the robotic device, testing if these measures collectively can predict outcome on the final day. We found in a cross-validation study that predictions are possible, yielding median r-squared values over 99%. Several particular measures were found to dominate their contribution to the prediction of recoverability. These results show that interactive self-rehabilitation may be a viable method for motor restoration, and the quantitative metrics available can be used to predict the eventual state of recovery.

I. Introduction

Despite evidence of possible recovery long after the onset of stroke [1], regaining functional use of the upper extremity can be elusive [2]. Emerging interventions including intensive repetitive practice [3, 4], task-specific training [5, 6], and interactive robotic technology [7–9] all aim to restore upper extremity motor ability and function. Although many of these studies focus on isolated limb actions, patients also care about completing the functional task with proper coordination of both arms [10]. While these methods offer significant benefits, many daily activities require a coordinated effort by both arms that might be more achievable through self-therapy.

A number of studies have investigated the efficacy of bimanual training on the recovery of the affected limb [11–15]. Others have stated that bimanual training engages additional cortical areas of the brain [16, 17]; hence it allows the possibility for additional “solo” training to occur after the one-on-one therapy has been exhausted. It remains to be tested whether such solo training might lead to added benefit to the therapy process if the proper technology is employed.

One reason why the prospects of bimanual training have not been fully understood is that it is a broad topic with many choices on the specific manner in which people practice. One can choose to move with hands physically coupled or uncoupled, in a mirror mode or in a parallel mode, with both hands moving together or in sequence. Previously, parallel reaching has been shown to imply less of a challenge in healthy individuals compared to reaching in a mirror mode [18, 19]. Our group has also investigated the possible advantage of simple simultaneous bimanual motion over practicing sequentially, when one arm performs an action after the other. Consequently, our work has shed light on the most likely successful mode for self-rehabilitation: simultaneous movements in parallel mode. However, it remains to be tested whether these healthy study behaviors would translate to the stroke population.

The manipulation of error signals during practice appears to stimulate improvement in coordination for individuals with or without a history of stroke [20]. In simple terms, if one perceives a larger mistake, they are motivated and naturally inclined to reduce the errors. Such error-driven learning processes are believed to be central to neuroplasticity and reacquisition of skill [21, 22], which leverages the natural adaptive nature of the nervous system [23].

While the mechanisms for these improvements are not yet known, based on our recent study [24] our group took error augmentation ideas and expanded them to the application of a therapist-patient-machine trio that works together to restore reaching ability. The therapist held a tracking device and provided a cue to the patient, which unbeknownst to them had their real-time errors magnified both visually and haptically (through robot-applied forces). Error augmentation demonstrated an advantage over and above repetitive practice alone. The obvious next question, however, is whether the cue might come from the patient’s unaffected arm, allowing for self-rehabilitation. Here, we expand on this novel concept to test a self-rehabilitation system that employs visual display and robotic technology with augmented error signals during training.

Furthermore, the most critical aspect such studies is the variance in improvement across the stroke population. So while we do test the efficacy of the treatment in a clinical design, we also seek to explore why some subjects might respond better than others. This may be inherent to each subject, and evidence may be present on the first day in the laboratory. There may be important clues from measures of motor performance and learning that can be used to construct such a prognostic model of recovery and reveal why some subjects respond to treatment better than others. The answers should come from cross-validated regression modeling, trying to relate all the data we have on each subject to their clinical progress during the study.

Accordingly, our aim in the present study was to train subjects in 7 visits over three weeks, and then to determine whether information available on their day 1 visit might be used to predict the gains seen across the three weeks. We were also interested in identifying movement metrics that may play a larger role in predicting those gains.

II. Methods

A. Subjects

Twenty-eight adults with chronic stroke agreed to participate in the study (11 Female, age range 26–78, mean age 55. 38). Study participants were recruited from a registry of post-stroke individuals or who contacted the lab with interest in participating due to postings in the Chicago area. Prior to commencing the study all participants provided informed consent according to the Declaration of Helsinki, approved University Human Subjects Internal Review Boards. Twenty-eight individuals began but two dropped out due to unrelated medical reasons and were excluded from analysis.

Eligible participants were all adults who had suffered a single cortical stroke and were at least six months post-stroke. Participation also required some recovery of proximal strength in the hemiparetic limb as confirmed by an upper extremity Fugl-Meyer score of 25–49. Exclusion criteria included multiple strokes, bilateral paresis, severe spasticity or contracture, severe concurrent medical problems, severe sensory deficits, cerebellar strokes resulting in severe ataxia, significant shoulder pain, focal tone management with Botulinim Toxin (Botox®) injection to the hemiparetic upper extremity within the previous four months, depth perception impairment (< 3/9 on Stereo Circle Test), visual field cut, cognitive impairment (Mini Mental State Examination < 23/30), or severe aphasia, affective dysfunction or hemisensory neglect that would influence the ability to perform the experiment or provide informed consent. Participants were also excluded if they were currently receiving any other skilled upper extremity rehabilitation in a clinical setting.

B. Study Setting

The study used a three-dimensional haptic/graphic system called the Virtual Reality Robotic and Optical Operations Machine (VRROOM) [25]. A cinema-quality digital projector (Christie Mirage 3000 DLP) displays the stereo images that span five-foot-wide 1280×1024 pixel display resulting in a 110° wide viewing angle in a see-through augmented reality display. In this study, vision of the arms was occluded so that only cursors (representing hand locations) and targets were shown. Infra-red emitters synchronize separate left and right eye images through StereoGraphics liquid crystal shutter glasses. An Ascension Flock of Birds sensor tracked head motion for appropriate display of perspective, while another sensor served as the position tracker of the non-affected wrist. A SensAble Technologies Phantom Premium 3.0 robot interfaced with the participant’s impaired wrist (Fig. 1). A Wilmington Robotic Exoskeleton (WREX^™) provided anti-gravity arm support [26].

Figure 1:

VRROOM Robotic bimanual tele-rehabilitation system. A tracking device on the less impaired limb was used to present a target cursor for the imparied limb to follow.

C. Protocol

Initial Fugl-Meyer scores from each participant were used to block-randomize into two groups – error augmentation (EA) and control. Each group received two weeks of training with participants receiving three, 45-minute sessions per week (six sessions total). After a week of rest, each participant went through a follow-up evaluation.

Each session began with six 5-minute blocks of movement training with two-minute rest periods between each treatment block. The treatment protocol included the practice of bimanual reaching movements for all participants in addition to free movements. The free movement blocks targeted participants’ idea of good therapy based on their need with the possibility of choosing the previous standardized five-minute block practice. This allowed the participants to customize their own therapeutic approach, focusing on what was most critical for their particular weakness areas. Quantitative assessments were performed at the beginning and end of the treatment (pre and post) as well as one week after the post assessment (follow-up).

During all therapy sessions, participants were comfortably seated in a chair with the hemiparetic arm supported by the WREX^™ gravity-balanced orthosis. The hemiparetic hand was placed in an exotendon glove that included a wrist splint, which assisted with hand opening and neutral wrist alignment to allow for a more functional hand and wrist position. Since holding a handle is not necessarily the same as free-hand motion [27], we connected the robot near the wrist to allow the hand to open freely as well as allow free pronation and supination of the forearm with the WREX^™ swiveling wrist support. Both the PHANTOM^™ robot and position tracker were attached to the affected and non-affected forearms respectively, with the center of the devices located above the radiocarpal joint. Forces were applied by the robot during the EA treatment; however, the robot was attached during both treatments to assist in blinding the participant as well as to provide feedback regarding location within the 3D workspace.

During training, participants were only able to see two cursors within the virtual environment with the view of their arms being blocked. Each cursor displayed the movement of one of their arms. Participants were instructed to keep moving their arms together as much as possible while reaching to targets throughout the workspace. For the EA treatment, the error vector e, defined as the instantaneous difference in position between the participant’s wrists was visually magnified by a factor of 1.5 as part of the error augmentation. Additionally, an error augmenting force of 100 N/m was applied pushing the participant’s affected hand further away from the non-affected hand. For safety purposes, this force was designed to saturate at 4 N.

D. Analyses

Participants were evaluated inside the VRROOM using their progress from one difficulty level to the next and outside the VRROOM with the clinical measures immediately prior to the 2-week treatment phase and again at the end of the treatment phase. Follow-up testing was performed one week after the end of treatment. A blinded evaluator administered all outcome measures including our primary outcome; the arm motor section of the Fugl-Meyer (AMFM) to measure impairments [28, 29] as well as our secondary outcome measures, which included the Wolf Motor Function Test (WMFT) for functional ability [30, 31], Motor Activity Log (MAL) for quality and quantity of arm use in activities of daily living [32, 33] and the Box and Blocks assessment as an indicator of manual dexterity [28, 34]. Finally, to assess perception of the experience, participants completed the Intrinsic Motivation Inventory (IMI) questionnaire [35], which consists of 25 questions in four categories (interest/enjoyment, perceived competence, motivation/effort, and perceived value).

We extracted sixteen features from each movement made by the patient, described in detail below, these features were averaged, and their maximum value was computed as well as their variance across the twenty evaluation reaches, yielding forty-eight total features. Demographic and evaluation information, was also included, bringing the total to sixty-four features.

Demographics:

Age, height, mass, gender (binary), time since stroke (in months), dominant arm (binary), affected arm (binary), dominant arm same as affected arm? (binary), stroke type (Ischemic vs. Hemorrhagic, binary), stroke location (divided into cortical, subcortical, and brainstem, all binary), treatment (EA vs. standard, binary)

Initial evaluation:

Initial Fugl-Meyer (AMFM) score, initial Wolf Motor Function (WMFT) score, initial Box and Blocks score

Movement metrics:

• Reaction Time: Time from the moment the target is shown until the subject begins to move (red triangle in Fig. 2A).

• Maximum Speed: The fastest speed (magnitude of velocity) the subject achieved during a reach (blue triangle, Fig. 2A).

• Hand path length: total distance traveled by the hand (cursor) in a reach (black trajectory in Fig. 2B)

• Path length ratio: ratio of the hand path length divided by the length of the straight line path to the target (length of black trajectory divided by length of blue trajectory in Fig. 2B).

• Initial Direction Error: angle between the initial trajectory leaving the starting position and the straight light trajectory to the target (Fig. 2C).

• Pre-movement speed: average speed from the moment the target is shown until the subject launches their movement (between t = 0 and t = reaction time)

• Trial time: time from when the target is first displayed until it is reached.

• Speed range: difference between maximum speed and minimum speed.

• Initial movement ratio: distance covered by the initial launch divided by the total distance traveled (hand path length)

• Speed ratio: ratio of the speed achieved during initial launch and max speed

  $$\mathit{\text{speed\hspace{0.17em}ratio}}=\frac{\mathit{\text{speed\hspace{0.17em}of\hspace{0.17em}initial\hspace{0.17em}movement}}}{\mathit{\text{maximum\hspace{0.17em}speed}}}$$

• Number of speed peaks: this metric is inversely related to movement smoothness. More peaks means the subject’s movement was more fragmented and contained more “sub-movements,”

• Maximum perpendicular distance: maximum deviation from the straight line path between the starting point and the target

• Time for max speed: time when maximum speed was achieved

• Time for max perpendicular distance: time when the subject was furthest from the straight line path

• Percentage of Movement in the Target Direction (PMTD): it is equal to the sum of the projections of the vector between each two time points onto the vector connecting the starting point and the target (orange line in Fig. 2D)

• Mean Arrest Period Ratio (MAPR): ratio of the time spent below a certain speed threshold and the total movement time (trial time)

  $$\mathit{\text{MAPR}}=\frac{\mathit{\text{time\hspace{0.17em}spent below the red line in}}\text{ Fig. 2}E}{\mathit{\text{trial time}}}$$

Figure 2.

Explanation of certain movement metrics. A: Reaction time (red triangle) and Maximum Speed (blue triangle). B: Path Length Ratio. C: Initial direction error (angle α). D: Percentage of movement in the target direction (PMTD). E: Mean Arrest Period Ratio (MAPR)

E. Statistical Analysis

To examine treatment-related change, outcomes were analyzed using a repeated measures analysis of variance (ANOVA), with factors of time (pre vs. post vs. follow-up) and treatment type (EA vs. standard). Finally, Tukey HSD post-hoc analysis was performed when necessary to evaluate detailed changes in participants’ performance. All statistical tests were evaluated using an alpha level of 0.05.

Since we have sixty-four features, it is extremely easy to overfit the data, hence we decided to reduce the dimensionality of the model by using principal components. For the cross-validation process, with twenty-six subjects, we decided to leave out ~8% (2 subjects) at a time, for a total of 325 possible combinations. We fit linear models, models that consider pairwise interactions between linear terms (interaction models), and quadratic models (containing both linear, pairwise interaction, and quadratic terms) using one, three, five, seven, nine, and eleven principal components. Our main outcome measure was the change in the patient’s AMFM. All models were cross-validated and their r² with the change in the patient’s AMFM score was computed.

Next, we wanted to determine which features were critical in contributing to the change in AMFM scores. We excluded one feature at a time from our dataset and ran the principal component analysis using all the other features, then we constructed a quadratic model in each case using 7 principal components.

III. Results

A. Clinical intervention

Our main outcome measure, AMFM, in an overall analysis showed significant improvement over the three weeks (F(2,36)=3.96, p<0.05), such that the average gain was 2.90 ± 5.24 points. Further detailed analysis failed to detect a significant difference between standard and EA treatments (Fig. 3). However, the group that received EA treatment did improve significantly from the pretreatment evaluation to the one-week follow-up with an average gain of 3.50 ± 3.47 (t=3.19, df=9, p<0.05). Interestingly, this group also improved significantly over the rest period with an average gain of 2.60 ± 3.50 (t=2.35, df=9, p<0.05)

Figure 3.

Change in Fugl-Meyer across evaluations, performed on each participant’s first visit, final treatment visit (visit 6) and 1 week follow-up (F/Up). Red represents error augmentation treatment while blue represents control.

A. Prognostic modeling results

The results of fitting the models with a different number of principal components are shown in Fig 4. All models were fit using a stepwise process that minimizes the mean square error and were cross-validated. From the figure, we can see the worst performance from linear models, and the best performance from the quadratic models. Both the interaction models and the quadratic models cross an r² of 0.95 with 7 principal components, linear models do not reach such high performance. The size of the shaded area (interquartile range, IQR) diminishes for both interaction models and quadratic models up to 7 principal components. Beyond that, the quadratic models’ IQR continues to shrink while the interaction models’ IQR increases.

Figure 4.

Model performance with a different number of principal components. The lines connect the medians of the model fits, and the shaded regions cover the interquartile range.

When we excluded one feature at a time from our analysis to understand the effect on AMFM prediction error, and after performing the 325-fold cross-validations in each case, we obtained the results shown in Fig. 5. From these results we can see that removing certain features increases AMFM prediction error, these features are: mean reaction time, mean MAPR, longest time for maximum perpendicular distance, maximum MAPR, variance of the patient’s reaction time, and variance of movement smoothness (variance in number of speed peaks). Excluding other features caused negative prediction error, yet only one did so significantly on the 95% confidence level: initial box & blocks score.

Figure 5.

Mean and 95% confidence interval for precdiction error when each feature is excluded from the model.

IV. Discussion

This blinded, randomized study revealed a benefit to arm recovery and functional use with the proposed self-rehabilitation system. The Fugl-Meyer score showed significant improvement over the three-week intervention. This study did not, however, establish a difference between EA and standard treatments. An additional key result was that a simple, subject-specific regression model can predict the functional gains using information from the initial day.

While gains in clinical scores were modest, they were not clinically meaningful [36]. However we argue that such gains might grow in longer treatment courses since in contrast to this study, interventions often train for six weeks or more [7, 37–39]. This may be also why we failed to find the significant advantage to EA treatment, which was observed in our previous study [24]. However, removing one outlier from the standard treatment group resulted in similar trends to our previous study. Interestingly, this particular subject was only six months post stroke, and may have been exhibiting the larger improvements more normally associated with sub-acute recovery.

Acceptable performance with Fugl-Meyer (r² of 0.95 and above) is achieved using a quadratic model and seven principal components. The three types of models (linear, pairwise interaction, and quadratic) have similar performance up to 3 principal components. Poor performance of the linear models suggests a higher order relationship between the principal components and the outcome measure – rather than a simple linear sum – which is confirmed by the better performance of the interaction and quadratic models. Medians and interquartile ranges were used instead of means and confidence intervals because the data is not normal, especially beyond 5 principal components, this might be the reason for the degradation in performance for the interaction models at 9 principal components: there is a wide spread of data points below the median, while the spread above the median is small, leading to a drop in the resulting models’ median r². The quadratic models cross 0.95 r² at 7 principal components, and saturate around 1 after that.

Excluding features from the analysis revealed that features related to disturbance/error response and movement smoothness are essential when trying to predict AMFM change, since excluding them from the analysis increased the error. Further analysis is required to evaluate whether excluding more than one feature, or “trimming down” the features being included in the principal components could have a more significant impact on reducing AMFM prediction error. We are also exploring the predictability of two other clinical scores: Wolf Motor Function Test (WMFT) and Box & Blocks.

This study provides practical evidence supporting self-rehabilitation, but more importantly it provides modeling analysis that supports the possibility of prognostication – that the degree of a patient’s 3-week outcome is predictable from data available only on the first day. Furthermore, the benefits associated with error augmentation may become clearer for larger dosages and longer durations of practice. The regression modeling results demonstrate through cross validation that a person’s gains can be predicted with very high accuracy. In the search for optimal training methods, the evidence presented here also sheds light on the important ability of quantitative measures to prognosticate using factors measured by the very same technologies used to apply therapy.