Predicting Improved Daily Use of the More Affected Arm Poststroke Following Constraint-Induced Movement Therapy

Abstract

Background

Constraint-induced movement therapy (CI therapy) produces, on average, large and clinically meaningful improvements in the daily use of a more affected upper extremity in individuals with hemiparesis. However, individual responses vary widely.

Objective

The study objective was to investigate the extent to which individual characteristics before treatment predict improved use of the more affected arm following CI therapy.

Design

This study was a retrospective analysis of 47 people who had chronic (> 6 months) mild to moderate upper extremity hemiparesis and were consecutively enrolled in 2 CI therapy randomized controlled trials.

Methods

An enhanced probabilistic neural network model predicted whether individuals showed a low, medium, or high response to CI therapy, as measured with the Motor Activity Log, on the basis of the following baseline assessments: Wolf Motor Function Test, Semmes-Weinstein Monofilament Test of touch threshold, Motor Activity Log, and Montreal Cognitive Assessment. Then, a neural dynamic classification algorithm was applied to improve prognostic accuracy using the most accurate combination obtained in the previous step.

Results

Motor ability and tactile sense predicted improvement in arm use for daily activities following intensive upper extremity rehabilitation with an accuracy of nearly 100%. Complex patterns of interaction among these predictors were observed.

Limitations

The fact that this study was a retrospective analysis with a moderate sample size was a limitation.

Conclusions

Advanced machine learning/classification algorithms produce more accurate personalized predictions of rehabilitation outcomes than commonly used general linear models.

Upper extremity hemiparesis occurs in up to 77% of stroke survivors.¹ It has a large effect on functional independence^2,3 and leads to increased caregiver and societal burden.^3–5 Many upper extremity rehabilitation interventions target motor functioning without addressing the daily use of that limb during functional activities outside the laboratory.⁶ However, spontaneous arm use during daily activities influences quality of life more than motor ability does,⁷ which therefore makes it an important consideration for patient-centered care.

Constraint-induced movement therapy (CI therapy) is one of just a few interventions known to promote improvements in daily arm use.⁶ It is a time-intensive intervention requiring extensive patient and therapist resources. Therefore, it should be undertaken when it is most likely that the patient will experience a clinically significant reversal of nonuse, thereby justifying the investment. As pretreatment motor ability has emerged as the primary predictor of change in daily use after CI therapy,^7,8 the current criterion used to identify appropriate candidates for CI therapy is presence of some active motion throughout the arm and hand.⁹ However, when using this criterion, individual treatment responses vary markedly.^10,11 This indicates that other factors influence recovery of arm use.

No studies to date have examined which elements of motor function are the most critical for achieving robust gains in arm use nor how other prevalent deficits that impede daily activities after stroke^12–15 (eg, sensation, cognition) may moderate this response.^16,17 As cognitive impairment has often been used as an exclusionary criterion in CI therapy trials,^18,19 its impact on learned nonuse remains particularly elusive, with existing evidence limited to underpowered negative findings.^16,20 By identifying how baseline participant characteristics (obtained via commonly used clinical outcome measures) influence improvements in arm use following CI therapy, therapists can empirically match the treatment approach to the client. Accordingly, health care resources could be used more efficiently to maximize patient outcomes.

To date, the insufficiency of current analytical approaches has been a barrier to realizing these efficiencies in patient care. This is because outcomes prediction using general linear models is too imprecise to be valuable for clinical decision making.^7,8 Although statistical regression techniques are computationally efficient, they are not able to capture complex nonlinear relationships among predictors in order to produce the most accurate prognostic models.^21,22 Novel computational modeling based on machine learning^23–27 can improve prognostic accuracy substantially (ie, the vast majority of people respond as predicted).^28,29 Recently, the enhanced probabilistic neural network model (EPNN) of Ahmadlou and Adeli³⁰ was used to predict extent of motor restoration after CI therapy with high accuracy.^31,32 This prior work, however, did not examine the extent to which patients in the chronic phase can continue to improve daily use of the affected side even in the absence of continued motor recovery.^30–32 As it is improved daily arm use—not recovery of motor function per se—that is associated with improvement in quality of life,³³ the analysis presented here extends prior work by predicting a person’s improvement in daily arm use after CI therapy.

Based on previous findings,^7,34 the authors hypothesized that better pretreatment motor ability would yield greater improvements in daily use of a hemiparetic arm. Additionally, given the impact of touch sensation and cognition on ability to complete activities of daily living,^13–15 better sensation and cognition were hypothesized to facilitate gains in arm use following CI therapy. The computational model previously developed by the authors³¹ and a neural dynamic classification algorithm developed by Rafiei and Adeli³⁵ quantitatively tested these hypotheses by identifying the combination of predictors that yielded the most accurate prognosis. A sensitivity analysis assessed the relative contribution of each predictor. Finally, visual inspection of interactions between the most influential predictors produced clinical interpretation from the quantitative results (ie, aided knowledge generation versus strict mathematic prediction).

Methods

Participants

A consecutive retrospective sample of 47 people with chronic (> 6 months) mild to moderate upper extremity hemiparesis had been randomized to receive CI therapy in 1 of 2 protocols (NCT01725919 and NCT02631850).⁹ Participants were enrolled regardless of cognitive or mobility status, but all met the minimum motor criteria in Table 1.⁹Table 2 presents the participant demographics and baseline clinical characteristics.

Table 1.

Minimum Active and Passive Range of Motion (ROM) Required for Participation^a

ROM	Shoulder	Elbow	Wrist	Fingers	Thumb
Minimum passive	Flexion ≥ 90° Abduction ≥ 90° External rotation ≥ 45°	Extension ≥ 150°	Extension ≥ 0° Forearm supination/pronation ≥ 45°	MCP extension ≥ 145°	Extension or abduction of thumb ≥ 10°
Minimum active	Flexion ≥ 45° Abduction ≥ 45°	Extension ≥ 20° from a 90° flexed starting position	Extension ≥ 10° from a fully flexed starting position	Extension of MCP (and PIP or DIP) joints of at least 2 fingers ≥ 10°	Extension or abduction of thumb ≥ 10°

^aDIP = distal interphalangeal; MCP = metacarpophalangeal; PIP = proximal interphalangeal.

Table 2.

Demographic and Baseline Clinical Characteristics by Study^a

	Study 1 (n = 29)				Study 2 (n = 18)
Characteristic	%	Mean	SD	Range	%	Mean	SD	Range	Between-Study Difference (P)
Age (y)		58.84	12.71	24.19 to 84.01		61.28	17.23	27.23 to 81.79	.597
% Men	75.9	N/A	N/A	N/A	72.2	N/A	N/A	N/A	N/A
% With affected side on right	58.6	N/A	N/A	N/A	61.2	N/A	N/A	N/A	N/A
Chronicity (y)		3.05	3.76	0.57 to 19.62		2.60	3.31	0.53 to 11.83	.697
WMFT fine motor score (mean time, s)		44.80	43.74	2.73 to 120.00		28.73	36.73	2.60 to 120.00	N/A
WMFT gross motor score (mean time, s)		9.18	22.17	0.87 to 120.00		2.40	2.63	0.53 to 10.23	N/A
Semmes-Weinstein Monofilament Test score (raw)		63.09	123.17	0.01 to 300.00		17.89	70.43	0.02 to 300.00	N/A
WMFT fine motor score (ln)		3.23	1.18	1.00 to 4.79		2.56	1.31	0.96 to 4.79	.075
WMFT gross motor score (ln)		1.26	1.18	−0.14 to 4.79		0.47	0.86	−0.63 to 2.33	.017b
Semmes-Weinstein Monofilament Test score (ln)		−0.03	3.45	−4.83 to 5.70		−0.43	2.12	−3.91 to 5.70	.663
Montreal Cognitive Assessment score (raw)		24.10	4.47	14.00 to 30.00		21.83	5.88	7.00 to 30.00	0.141
Motor Activity Log score (raw)		1.20	0.85	0.00 to 3.04		1.61	0.94	0.15 to 3.20	.123

^aIndependent t tests were used for continuous variables, and chi-squared tests were used for categorical variables. Between-study differences were considered significant at P ≤ .05. n = natural logarithm; N/A = not applicable; WMFT = Wolf Motor Function Test.

^bSignificant difference.

Interventions

CI therapy comprised 30 hours (3 h/d, 10 weekdays, over a 3-week period) of intensive motor training focusing on functional tasks and shaping, plus behavioral techniques (Transfer Package), for 0.5 hour per day (5 hours total) with the aim of increasing the use of the affected limb during everyday activities.³⁶ Transfer Package techniques included the following: a treatment contract that collaboratively established behavioral targets for use of the more affected upper extremity, daily monitoring of arm use for daily activities through the Motor Activity Log (MAL), and problem solving through barriers to using the weaker arm for daily activities.^9,36 Participants in both protocols were encouraged to wear a padded mitt restraint on their less affected hand during waking hours to discourage using their less affected hand. Participants were strongly encouraged to use their more affected hand exclusively to perform daily activities whenever this was safe and possible. All treatments were delivered or supervised by a licensed physical or occupational therapist.

Assessment

The following motor, cognitive, and sensory assessments were administered by a masked assessor within 1 week prior to and following completion of CI therapy.

Motor Activity Log (MAL)

The MAL is a commonly used, reliable, and valid measure of daily use of the more affected arm.³⁷ The quality of movement of 28 daily activities was rated by participants on a scale that ranges from 0 (not attempted) to 5 (normal movement) with half-point increments. The total score for quality of movement reflects the mean of the 28 item scores. The amount of use scale was not used because prior research has shown redundancy of this scale with the quality of movement scale. The MAL has a minimal detectable change of 0.5³⁸ and a minimal clinically important difference of 1.0.³⁹

Wolf Motor Function Test

The Wolf Motor Function Test is a valid and reliable in-laboratory measure of upper extremity motor performance containing 15 timed functional tasks.^40,41 Following a factor analysis that revealed superiority of a 2-factor model to a 1-factor model, tasks were subdivided by whether they loaded primarily onto a gross motor function factor or a fine motor function factor, consistent with George et al.³² The 8 primarily fine motor tasks consisted of the following: lift can, lift pencil, lift paper clip, stack checkers, flip card, turn key in lock, fold towel, and lift basket. The 7 primarily gross motor tasks consisted of the following: forearm to table, forearm to box, extend elbow, hand to table, hand to box, weight to box, and reach and retrieve. The performance time of each item was natural logarithm transformed to reflect the relative nonlinearity of potential performance time improvement (ie, an improvement from 4 seconds to 2 seconds is greater than an improvement from 102 seconds to 100 seconds). The Wolf Motor Function Test fine and gross motor summary scores reflect the mean of the natural logarithm–transformed item scores; higher scores reflect poorer performance.^31,32

Semmes-Weinstein Monofilament Test

The Semmes-Weinstein Monofilament Test is a measure of touch threshold with good psychometric properties.^24,42,43 Without visual input, participants report presence or absence of touch sensation when the test administrator touches a monofilament to the surface of the skin from a perpendicular angle until the fiber buckles.⁴³ Grams of pressure required to consistently demonstrate index fingertip touch perception are recorded. Higher scores reflect more impaired sensation. Touch sensation was natural logarithm transformed because of positive skew.

Montreal Cognitive Assessment

The Montreal Cognitive Assessment is a cognitive screening tool that grossly assesses short-term memory recall, visuospatial abilities, executive functioning, language, orientation, attention, concentration, and working memory.^44,45 It is a valid and reliable instrument for estimating level of cognitive impairment within the stroke population,^44,45 with better sensitivity than the Mini-Mental State Examination.⁴⁶ Scores range from 0 to 30, with a traditional cutoff of 26 indicating the presence of mild cognitive impairment (higher scores are better).⁴⁷

Prognostic Model

Outputs

Three output categories were defined to describe the improvement in arm use after the intervention (MAL score change from before therapy to after therapy): (1) poor responder when the MAL change was < 1.0 (failure to reach minimal clinically important difference), (2) moderate responder when the MAL change was between 1.0 and 2.0, and (3) best responder when the MAL change was > 2.0. These category thresholds are visualized in the histogram of MAL score changes provided in the Supplementary Figure (available at https://academic.oup.com/ptj).

Predictors

The data from this analysis were obtained from 2 separate experiments (with identical inclusion criteria and recruitment strategy), and thus study sample (study 1 or study 2) served as a binary nuisance covariate/predictor (variable unrelated to the study question that may moderate the effects of the other variables); this controlled for potential selection factors and differences in personnel delivering the intervention. Other predictors included fine and gross motor ability before treatment, as measured by the Wolf Motor Function Test performance time, touch sensation (Semmes-Weinstein Monofilament Test), cognitive function (Montreal Cognitive Assessment), and pretreatment daily arm use (MAL). Raw data from the 6 independent predictors is provided in the Supplementary Table (available at https://academic.oup.com/ptj). Using the traditional Montreal Cognitive Assessment cutoff of 26 of 30,⁴⁷ 28 of 47 participants were categorized as having cognitive impairment. Side of stroke and concordance (ie, the affected side is also the dominant side) were not included in the model because of a general consensus in the literature that they are not important predictors of arm use after stroke.^17,34 Among 282 values in the original database, 3 were missing (1%). They were replaced using linear regression interpolation.

Computational model

Classification is a supervised machine learning approach that can be used to predict an individual outcome (categorically) based on prior data.^48–50 To maximize the probability of correct classification, Ahmadlou and Adeli³⁰ used local decision hyperspheres to modify the Gaussian function in probabilistic neural networks. The Gaussian function calculates the likelihood (probability) that a data point belongs to each outcome category, whereas the local decision hyperspheres serve to spatially distort the variable space based on the distribution of data points. By integrating local decision hyperspheres into the Gaussian model, one can account for complex and nonlinear relationships of the predictor variables on arm use. The end result is that an individual with clinical characteristics that are more similar to others who are known poor responders would receive a poor prognosis (predicted poor responder). EPNN has yielded accurate results for computer-aided diagnosis/prognosis of neurologic conditions such as Parkinson disease,⁵¹ mild cognitive impairment,⁵² major depressive disorder,⁵³ and autism spectrum disorder.⁵⁴

The prognosis model evaluated all possible combinations of 6 inputs using EPNN. For each combination, 47 prognoses were performed, each time selecting 1 participant’s data as test and the rest as training data for EPNN (N = 47; hence, this was performed 47 times per combination using a leave-one-out approach). EPNN assigned the test participant to an output/response class based on the similarities of his/her pretreatment characteristics to those of other (training) participants for each response class. The accuracy associated with each combination is the average accuracy of 47 prognoses. Details of the prognosis computational model can be found in George et al.³¹

Once the combinations with maximum accuracy were identified through EPNN, a neural dynamic classification algorithm (NDC) was used to investigate if a more accurate prognosis was obtainable. NDC uses the patented neural dynamic optimization developed by Adeli and Park⁵⁵ (NDAP; US Patent Number: 5,815,394) to discover a feature space with marginalized clusters and proximity of classmates.

The NDC consists of 5 layers: input, feature vector representor, pattern, summation, and decision. In the first layer, a selected combination of inputs was fed into NDC. In the feature vector representor layer, the inputs were transformed into another Euclidean space with a different dimension using 3 optimized linear transformation functions corresponding to poor, moderate, and best responders. These optimized functions minimize the proximity of classmates and marginalize nonclassmates in the new Euclidean space. These functions are optimized iteratively by NDAP. Pattern, summation, and decision layers are similar to those of EPNN. More details about NDC can be found in Rafiei and Adeli.³⁵

Role of the Funding Source

Financial support for data analysis was obtained through The Ohio State University Office of the Provost Chronic Brain Injury Discovery Theme initiative, the American Heart Association, and the Patient-Centered Outcomes Research Institute (PCORI). Additional support for participant recruitment and regulatory affairs was provided by the Center for Clinical and Translational Sciences (National Center for Advancing Translational Sciences, Grant 8UL1TR000090–05). The funders played no role in the design, conduct, or reporting of this study. The statements in this publication are solely the responsibility of the authors and do not necessarily represent the views of the funding agencies or their leadership.

Results

Enhanced Probabilistic Neural Network Results

Figure 1 presents the average rates of selection of 6 inputs in combinations with an accuracy of at least 75%. Based upon the divergence of predictors at around 85% accuracy, the decision was made to look at combinations at this level and higher. Table 3 presents the 9 combinations of predictors with an accuracy of at least 85%. The maximum accuracy of 91.5% was obtained by only 1 combination: study, fine motor, gross motor, and sensation (Tab. 3). Accuracies obtained using EPNN were substantially higher than the optimal general linear models using the same combinations of predictors (including interactions) (Tab. 3). Open-source code for applying the models in Table 3 to new data can be found at https://drive.google.com/open?id=1vOjw8JLgzP-BZ77067PPROeGPguwi4IX.

Figure 1.

Average rates of selection of 6 inputs in combinations with an accuracy of at least 75%.

Table 3.

Combination of Predictors Resulting in an Average Prognosis Accuracy of 85% or More^a

Combination	Study	Fine Motor Score	Gross Motor Score	Sensation	Cognition	Arm Use	% Accuracy of the GLM	% Accuracy of the EPNN
1	1	1	1	1	0	0	51.1	91.5
2	1	0	0	1	0	1	44.7	87.2
3	1	0	1	1	1	0	40.4	87.2
4	1	0	0	1	1	1	48.9	87.2
5	1	1	0	1	0	0	40.4	85.1
6	1	0	1	1	0	0	40.4	85.1
7	1	1	0	1	0	1	44.7	85.1
8	0	1	1	1	1	0	51.1	85.1
9	1	0	1	1	1	1	48.9	85.1
No. of times selected	8	4	5	9	4	4	N/A	N/A
% Selection rate	88.9	44.4	55.6	100.0	44.4	44.4	N/A	N/A

^aDesignated as 1 when the predictor was selected and 0 when it was not. EPNN = enhanced probabilistic neural network model; GLM = general linear model; N/A = not applicable.

Sensitivity Analysis of the Most Accurate Combination

To determine the relative importance of each of the predictors to the strength of the overall prediction, a sensitivity analysis was performed by removing predictors 1 at a time from the most accurate model in Table 3. When a predictor is more influential, the overall accuracy of the model will drop substantially following its removal. The accuracy of the model dropped from 91.5% (all predictors) to 83.0% (study removed), 85.1% (fine motor removed), 85.1% (gross motor removed), and 80.9% (sensation removed). These data indicate that sensation is more influential than the other predictors.

Final Improved Prognosis by Neural Dynamic Classification

NDC was used to improve the accuracy of the best combination (Tab. 3). Study, fine motor, gross motor, and sensation were inputted to the first layer of NDC. An improved prognosis accuracy of nearly 100% was achieved after 50 optimization iterations.

Clinical Insights Garnered From Visual Inspection of Scatterplots

Classification algorithms cannot indicate direction of effect. Therefore, to enhance the clinical relevance of findings from this research, a scatterplot was created to identify clusters of individual characteristics that appeared to be associated with poor or best treatment response among the predictors identified within the most accurate prognostic model. The scatterplot in Figure 2 depicts 3 main trends observed by the authors following visual inspection, also presented in Table 4: (1) those with the poorest sensation were moderate or best responders; (2) those who had both relatively intact sensation and motor ability were moderate or best responders; and (3) in the presence of relatively intact sensation, those with the poorest gross motor ability were poor responders.

Figure 2.

Scatterplot of the predictors in the most accurate prediction model. Poor responders improved less than the minimal clinically important difference (1.0); moderate responders’ Motor Activity Log (MAL) change was between 1.0 and 2.0; best responders had a MAL change of > 2.0. The main trends were as follows: (1) those with the poorest sensation were moderate or best responders; (2) those with better sensation and motor ability were moderate or best responders; (3) in the presence of relatively intact sensation, those with the poorest gross motor ability were poor responders. Although the spatial clustering of data points affects the classification, machine learning discovers spatial distortions in the variable space that cause some points to be statistically “closer” or farther from each other; such distortions would not be evident from this scatterplot. This 3-dimensional Figure is accessible for viewing via MATLAB at https://drive.google.com/file/d/14-pmb4ydT6zbXVqNZ_m_-ZLe2rZpFcgf/view?usp=sharing. WMFT = Wolf Motor Function Test.

Table 4.

Summary of Trends in Figure 2^a

	Predictor Variables
Figure 2 Trend	Sensation	Gross Motor Ability	Fine Motor Ability	Extent of Response
1b	Poorest	N/A	N/A	Best/moderate
2c	Better	Better	Better	Best/moderate
3d	Better	Poorest	N/A	Poor

^aN/A = not applicable.

^bThose with the poorest sensation were moderate or best responders.

^cThose with better sensation and motor ability were moderate or best responders.

^dIn the presence of relatively intact sensation, those with the poorest gross motor ability were poor responders.

Discussion

This study was designed to investigate clinical factors that can best predict the extent of improvement in daily use of a hemiparetic arm after CI therapy interventions. The most accurate combination at 91.5% (100% following NDC) included pretreatment gross motor ability, fine motor ability, and touch sensation, as well as study (nuisance covariate). Partially contrary to the hypothesis, baseline sensation, cognition, and motor ability failed to directly predict improvements in arm use following CI therapy: each of the models with high accuracy (> 85%) included 3, 4, or 5 predictors, and conversely, none of the models with a single predictor showed high accuracy. Rather, complex, nonlinear interactions among these factors accounted for the predictive accuracy of the model, as evidenced by high accuracies derived from the computational model and very low accuracies obtained from the (most accurate) general linear model.

Overall, touch sensation was the most influential predictor, as evidenced by its inclusion in every highly accurate model (ie, 100% selection rate in Tab. 3) and greater influence during sensitivity analysis of the top performing model. Moderate and best responders included those for whom sensation was nearly absent. Although this finding may appear to contradict previous findings that sensory loss associates with worse motor and functional outcomes,^56,57 the outcome measures used in these studies did not assess arm use, which is the main outcome being measured here. Arm use has been found to be a separate construct from motor function,³³ which limits comparison of these previous findings to the current study. Furthermore, in the few studies that have looked at predictors of arm use after CI therapy, none of them included touch sensation in their multivariate analyses.^7,14,16 Therefore, the findings presented here represent new information that could influence treatment decisions for people with mild to moderate chronic stroke hemiparesis.

Those with poor sensation may be better responders because motor function is available for enhancing arm use but prior use was dampened by lack of sensory input.^14,58,59 Improvement in performance of activities of daily living continues through experience-dependent learning even when improvements in neurologic impairments have slowed.^56,60,61 However, the initial awkwardness of using a limb with reduced sensation often results in nonuse of the affected arm. This phenomenon has been documented in monkeys with a deafferented limb⁶² and in those with sensory deficits after stroke.¹³ The CI therapy Transfer Package focuses on bringing attention to use of the affected arm throughout a person’s day,³⁶ which may be especially beneficial for those individuals who have sensory loss.

Among those with relatively intact sensation, motor ability predicted extent of improvement in arm use. This is consistent with previous findings that motor ability predicted improvements in arm use after CI therapy.^7,16 In these cases, recovery of motor ability may have been facilitated by better sensation prior to participating in the study, leaving a mainly behavioral deficit. According to the learned nonuse formulation, this deficit can be more easily overcome by the CI therapy Transfer Package because there are fewer barriers to successfully using the weaker arm for daily activities.⁶² Poorer gross motor ability (indicative of worse motor ability overall) appeared to be a barrier to improving arm use, whereas those with better motor function (fine and gross motor) made larger gains in arm use following CI therapy. Consistent with the hypothesis, this indicates that better motor function increases the likelihood of recovering substantial use of the weaker arm for daily activities when nonuse cannot be attributed to poor sensation. Though both gross and fine motor ability were influential in predicting recovery of daily use of a hemiparetic arm, gross motor ability was selected at a slightly higher rate in the most accurate models and it contributed more to accuracy as a single predictor (ie, 83.0% accuracy in a model including only study [nuisance covariate] and gross motor performance versus 78.7% accuracy in a model including only study and fine motor performance). The authors suspect that this is a result of gross motor function serving as a prerequisite for the utility of fine motor function. Despite the need for dexterous fine motor ability to complete many daily activities,⁶³ their initiation may not be possible without gross motor movements of the shoulder and elbow.⁶⁴ Therefore, gross motor ability may be a necessary precursor to arm use during daily activities.

Previous evidence has shown a strong relationship between improvement on the MAL and the inclusion of the Transfer Package.^36,65 The CI therapy Transfer Package involves extensive problem-solving and limb monitoring in the home environment with the need for recall of performance during therapy sessions.³⁶ Accordingly, it was initially hypothesized that a certain level of cognitive ability would be necessary in to order maximize the benefits of these methods.^66,67 Contrary to this hypothesis, cognition was not among the factors identified in the most accurate prognostic model, yet it was selected in other high accuracy combinations. This inconsistency may be accounted for by unmeasured interactions between level of family support and cognitive ability. For example, the adverse effects of decreased cognition on behavior change may be attenuated by increased behavioral reinforcement within a strong support system.^68,69 Ongoing work by the authors will lead to better understanding of the relationship between cognitive ability, family support, and effectiveness of the Transfer Package.

Finally, study (nuisance variable) was an influential predictor in the most accurate model, suggesting subtle differences in either study sample or treatment delivery (eg, different therapists) between the 2 otherwise identical protocols. Baseline gross motor function was worse for study 1 (Tab. 2), which, given that gross motor function influenced after treatment arm use, may partially account for participants in study 2 experiencing greater mean improvement on the MAL than those in study 1 (1.85 versus 1.32, respectively; P = .025 for significant time × group interaction in mixed-effects analysis of variance). Therapist skill in delivering the Transfer Package (or other unmeasured therapeutic or selection factors) may have also influenced treatment response. Indeed, provider characteristics, system dynamics, and techniques used to build a therapeutic alliance have all been shown to impact outcomes.^65,70–72 The finding that situational differences had an unanticipated effect on participant outcomes (confounding variable) reaffirmed the authors initial decision to include a nuisance covariate when combining datasets from otherwise identical treatment protocols, as failure to do so would otherwise have reduced the accuracy and/or validity of the predictive model.

Limitations of this study include a moderate sample size of individuals that were consecutively enrolled into 2 different prospective study protocols. Separating motor function into fine motor and gross motor performance, though aiding interpretation, also represents an oversimplification for 2 reasons. First, performance of fine motor tasks requires at least some proximal muscle movement. Second, Factor Analysis reveals that further subdividing these 2 major components (eg, splitting the gross motor factor into items that primarily involve elbow flexion/extension and those that do not) can explain an even larger portion of the variance in the larger construct of upper extremity ability. However, added complexity requires increased computational resources and reduces interpretability of the findings. This is because the classification algorithms used here are essentially “black boxes” of prediction whereby a clinician/researcher can input data from an individual and generate a prediction of treatment response. Knowledge generation thus requires the additional step of visualizing the data to understand the higher-order nonlinear interactions captured by this approach. Despite these limitations, the magnitude of the prognostic accuracy reported here supports the strength of the findings. Future directions will include replication with a larger sample, as well as examination of the influence of additional factors, such as extent of family support, to further improve prognostic accuracy.

Conclusion

It has been established that increased motor ability does not necessarily lead to increased use of a weaker arm.^38,73,74 In this paper the authors used novel computational methods to examine high dimensional nonlinear relationships among baseline characteristics to predict how well an individual could overcome nonuse of the weaker arm following CI therapy. Complex patterns not identifiable using traditional statistical regression models emerged. One group of best responders presented with poor sensation like the monkeys with a deafferented limb in the study of Taub et al.⁶² Another group of best responders presented with a pure behavioral deficit (learned nonuse) in the context of relatively intact sensorimotor function. Worst responders tended to present with the poorest motor function; effortful movement creates a continuous barrier to daily use of the hemiparetic arm. Taken together, these findings indicate that the CI Therapy Transfer Package is a particularly fruitful use of therapy time for individuals with sensory impairment and/or mild motor impairment who may be discharged prematurely from therapy after achieving a plateau in motor function. This ability to match an intervention to patient characteristics should facilitate more efficient use of health care resources and lead to more timely and improved patient outcomes.

Author Contributions

Concept/idea/research design: M.H. Rafiei, H. Adeli, L.V. Gauthier

Writing: M.H. Rafiei, K.M. Kelly, H. Adeli, L.V. Gauthier

Data collection: K.M. Kelly, A.L. Borstad

Data analysis: M.H. Rafiei, K.M. Kelly, L.V. Gauthier

Project management: K.M. Kelly, L.V. Gauthier

Fund procurement: L.V. Gauthier

Providing facilities/equipment: A.L. Borstad

Consultation (including review of manuscript before submitting): A.L. Borstad, H. Adeli

Ethics Approval

This study was approved by the Institutional Review Board at the Ohio State University. Every participant provided informed consent.

Funding

Financial support for data analysis was obtained through The Ohio State University Office of the Provost Chronic Brain Injury Discovery Theme initiative, the American Heart Association, and the Patient-Centered Outcomes Research Institute (PCORI). Additional support for participant recruitment and regulatory affairs was provided by the Center for Clinical and Translational Sciences (National Center for Advancing Translational Sciences, Grant 8UL1TR000090–05). The statements in this publication are solely the responsibility of the authors and do not necessarily represent the views of the funding agencies or their leadership.

Clinical Trial Registration

The trials were registered at ClinicalTrials.gov (NCT02631850 and NCT01725919).

Disclosure

The authors completed the ICMJE Form for Disclosure of Potential Conflicts of Interest and reported no conflicts of interest.

M.H. Rafiei, PhD, Whiting School of Engineering, Johns Hopkins University, Baltimore, Maryland.

K.M. Kelly, PT, DPT, EdM, Department of Neurology, The Ohio State University, Columbus, Ohio.

A.L. Borstad, PT, PhD, Department of Physical Therapy, The College of St Scholastica, Duluth, Minnesota.

H. Adeli, PhD, Department of Biomedical Informatics, Department of Neurology, Department of Neuroscience, The Ohio State University.

L.V. Gauthier, PhD, Department of Physical Therapy and Kinesiology, University of Massachusetts Lowell, 3 Solomon Way, Weed Hall 218D, Lowell, MA 01854 (USA).

Dr Kelly is a board-certified clinical specialist in neurologic physical therapy, a certified personal trainer, and a performance enhancement specialist.