PubMed, Pedro, and Web of Science databases were searched for studies published between January 2000 and March 2016. For PubMed and Web of Science databases the following key words were searched: (1) learning OR motor learning OR motor control AND (2) stroke OR cerebral palsy AND (3) parameter OR measure OR assessment OR adaptation OR acquisition OR retention OR transfer AND (4) upper limb OR upper extremity OR arm OR hand AND (5) rehabilitation OR treatment OR training. Due to the considerable number of key words, modification of the search strategy was made for the Pedro database, to match the database requirements. The following searches were performed: (1) cerebral palsy AND learning AND hand (2) cerebral palsy AND motor control and hand (3) stroke AND learning AND hand (4) stroke AND motor control AND hand. The key word “hand” was chosen to represent the upper limb studies, as from overview of the retrieved records, it included results of most studies performed on the upper extremity.

To be included in the review, a study had to: (1) involve either stroke survivors or persons with CP, (2) assess process or type of learning after an intervention with motor learning principles, (3) be applicable for rehabilitation, (4) use outcome measures relevant for motor control, (5) focus on the upper extremity, (6) be written in or translated to English. Studies were excluded from the review if the study: (1) did not focus on human subjects, (2) used only clinical outcome measures, (3) included an intervention of virtual reality, transcranial magnetic stimulation or of electric muscle stimulation.

Titles and abstracts of retrieved records were merged into one database on the reference management software, and duplicates were removed.

Titles and potential abstracts were screened independently by two researchers (First and last authors). Titles that contained any of the exclusion criteria were excluded based on title only. Relevant full text articles and full texts of abstracts that were inconclusive regarding their relevancy were assessed, and studies that did not correspond with the inclusion criteria were excluded. Fitting articles were also extracted from reviews relevant to the topic and from full text article references. Data regarding studies' designs was extracted. All study designs of any methodological quality were included. Due to our objective to perform a comprehensive data collection of the various parameters and measures, we did not factor the strength of experimental evidence provided by the studies. In addition to studies that examined the efficacy of an intervention, we included studies that explored the feasibility of tools, hypotheses regarding mechanisms of learning and recovery and the implementation of mathematical models. In such studies, assessment of the methodological quality would yield no benefit, due to their different objectives. A narrative synthesis of the literature was performed.

First, all included articles were reviewed. Data was extracted and compiled from each research on: (1) studies' characteristics and methodology, (2) the grouping method, (3) type of intervention, (4) study protocol (5) outcome measures. We further gathered information regarding the characteristics of the outcome measures utilized to assess learning. Parameters utilized to infer learning were clustered according to the motor learning principle or form of learning that they assessed. Underlying measures for each parameter were defined as the measures used to quantify the parameter.

Six of the reviewed studies examined the adaptation process (Dancause et al., 2002; Takahashi and Reinkensmeyer, 2003; Patton et al., 2006; Scheidt and Stoeckmann, 2007; Masia et al., 2011; Bourke et al., 2015). Two studies assessed the predictive control (Raghavan et al., 2006; Mawase et al., 2011).

Patton et al. (2006) and Takahashi and Reinkensmeyer (2003) evaluated stroke subjects performing robotic reaching movement training. A baseline phase without perturbations was followed by a learning phase with constant exposure to forces. Then, after-effects were measured by a catch trial phase that included intermittent removal of the force field. Finally, when forces were completely removed (i.e., washout period), de-adaptation was assessed (i.e., tendency of the after-effects to disappear). Patton et al. (2006) measured the initial direction error, testing the effectiveness of the shift of the early part of the movement, and the average shift in initial direction from the un-perturbed baseline trials to the after effects catch trials to establish the adaptation capacity. Takahashi and Reinkensmeyer (2003) also assessed the adaptation capacity of stroke subjects by the spatial reaching error from baseline. But the late correction (the distance from the maximal and final deviation from the reference path) was also measured, to assess the whole movement following applied and removed force fields. Performance improvement was assessed as the reduction in reaching error from baseline to final phase (Takahashi and Reinkensmeyer, 2003; Patton et al., 2006). Scheidt and Stoeckmann (2007) examined the adaptation to velocity dependent perturbations with pseudo-random magnitude, using measurement of the initial direction error, and the end point accuracy. The compensatory response acquired during training was measured by the hand movement onset speed, peak speed point, penultimate position point (i.e., the moment the speed dropped below 20% of its maximal value) and by the final position point. Dancause et al. (2002) explored error correction strategies following an unexpected spring-like load presented in 30% of elbow flexion movement trials. The strategies were identified by comparison between the angular positions and torques of the initial movement before the load was applied and the correction following the load. Bourke et al. (2015) assessed the corrective responses after unexpected external perturbations to the elbow or shoulder, while instructed to maintain their hand at a spatial goal. The response was quantified by the posture speed before the perturbation, deceleration time and maximal displacement after perturbation, return time to baseline position, end point error and by the joint velocity offset that represents the multi-joint coordination. Masia et al. (2011) explored the ability of children with CP to adapt to randomized center-out movements while performing reaching movements toward peripheral targets. For each reaching movement lateral deviation from straight line, acceleration peak, and peak and average speeds were measured at familiarization, force field adaptation, and washout phases. Directional analysis was performed for all measures to assess the anisotropy index (i.e., the roundness or flatness of the ellipse). The learning index was established as the degree of adaptation measured by the lateral deviations in force field and catch trials.

Mawase et al. (2011) explored the predictive control of CP subjects using a grasp and lift task of a virtual object. A sequence of increasing weights appeared randomly within trials of random weights. The planning of grasp precision was assessed by measurement of the grip force at the beginning of the lifting task and by the vertical trajectory of the object estimating the motor command. The precision of grasp execution was assessed by measurement of the temporal coordination. Raghavan et al. (2006) also explored the planning of precision grasp and precision of grasp execution among stroke patients. The peak grip and load forces and the timing and efficacy of grip load force coordination were measured respectively.

Chen et al. (2012) examined the reaching performance of children with CP by measurement of reaction time (RT), normalized movement time (NMT), normalized number of movement units (NMU) and the peak velocity (PV). Three studies assessed the performance of stroke patients performing a reaching movement task following an intervention (Wu et al., 2007; Caimmi et al., 2008; Durham et al., 2014). Wu et al. (2007) measured the RT, PV, movement time (MT) and the total displacement. Caimmi et al. (2008) measured the movement duration, end of movement angle, mean angular, and target approaching velocities, consistency of the target approaching and the movement smoothness (MS). The study objectives were to examine whether kinematic analysis is a sensitive and reliable measure, whether it can identify the mechanisms leading to improvement and quantify the functional improvement. Durham et al. (2014) measured the movement duration, PV, percentage time to peak velocity (%TPV), or to peak deceleration and to peak aperture, peak aperture size, peak elbow extension, and the MS. Wu et al. (2011) assessed the performance of stroke patients during pressing a desk bell and pulling a drawer tasks by measurement of the NMT, NMU, PV, percentage of movement time where peak velocity occurred, and the MS. Christopher and Johnson (2014) examined the performance of stroke patients performing a drinking task by measuring time to completion and the MS. Aluru et al. (2014) evaluated the performance of stroke subjects by measuring the movement speed, electromyographic activity of wrist extension, activation of wrist extensor and flexor, and co-activation of antagonist muscles during a wrist flexion-extension task.

While Wu et al. (2007), Aluru et al. (2014), Caimmi et al. (2008), Christopher and Johnson (2014), and Durham et al. (2014) studied only the performance, which is an individual execution of a skill (Krakauer and Mazzoni, 2011; Kitago and Krakauer, 2013), four studies (Thaut et al., 2002; Colombo et al., 2008; Massie et al., 2009; Geerdink et al., 2013) explored the skill acquisition, which depends on extended practice of the performance (Krakauer and Mazzoni, 2011; Kitago and Krakauer, 2013). Colombo et al. (2008), Thaut et al. (2002) assessed the performance change of stroke patients' reaching movements. Colombo et al. (2008), with the purpose of better understanding the learning mechanisms of stroke patients, measured the efficacy [the active movement index (AMI)], accuracy [mean distance (MD) from theoretical path], efficiency [normalized path length (nPL)], MS, MT, and the force control (error in orientation of force generation) of reaching movements. Massie et al. (2009) measured the trajectory variability, MT, reach velocity of reaching movements and the difference in compensatory strategies during treatment. Thaut et al. (2002) measured the movement durations, arm kinematics, variability of timing, reaching trajectories, and rhythmic synchronization. Geerdink et al. (2013) assessed the learning curve of manual dexterity of children with CP to establish the time during the intervention when maximal effects were reached using the box and block test (BBT) (counts the number of blocks that are transferred with a single hand from one compartment to another within 60 s).

Retention and transfer tests vary in the information received regarding the obtained learning. The transfer test assess a skill that was not practiced, whereas the retention test examines the trained task after a time interval (Schmidt and Lee, 2004; Kantak and Winstein, 2012).

Six studies explored whether trained movement generalized to untrained movement using transfer tests (Dipietro et al., 2007, 2009; Senesac et al., 2010; Dipietro et al., 2012; Krebs et al., 2012; Kitago et al., 2015). Dipietro et al. (2007, 2009, 2012), and Krebs et al. (2012) examined the generalization of trained reaching movements to untrained circle drawing movements by measurement of the axes ratio metric (i.e., indication for shoulder-elbow coordination). Additionally the studies measured speed profiles, MS, and measures of submovements [i.e., discrete ballistic movements that are a part of a more complex movement (Rohrer et al., 2004)] (not all measures were used in all studies, see Table 2). Dipietro et al. (2007) also examined changes in flexor-extensor abnormal synergies by measuring the joint angles correlation metric (independence of elbow and shoulder movements), orientation (best fitting line of hand path) and the major and minor axes of the drawn ellipse. Senesac et al. (2010) assessed the spatial generalization of improved proximal inter-joint coordination to two untrained reaching tasks, one spatially similar and the other different. Reach end-point kinematics (hand path curvature, time to peak velocity, PV, MS, and acceleration) were measured. Kitago et al. (2015) measured the reaching trajectories and their quality (i.e., trajectory analysis), MT, directional error, MS and and-point accuracy to assess the transfer of trained goal-directed reaching movements to untrained out-and-back straight movements. Two studies suggested that motor recovery after stroke and motor habilitation of children with CP resembles a motor learning model more than an adaptation model, as the trained movements generalized to the untrained movements (Dipietro et al., 2012; Krebs et al., 2012).

While previously discussed studies implemented transfer tests assessing the generalization from trained to untrained movements (Dipietro et al., 2007, 2009, 2012; Senesac et al., 2010; Krebs et al., 2012; Kitago et al., 2015), Sterpi et al. (2012) implemented a transfer test that assessed the generalization of movements trained in a certain workspace (reaching movements to form a path of a square) to a different workspace (within and outside the square) of stroke subjects. The AMI, MT, MD (error of movement accuracy), nPL (error of movement efficiency) and the MS were measured. Gilliaux et al. (2015) examined the performance of children with CP by measuring the amplitude and coefficient of variation (CV) of straightness for a reaching-as-far-as-possible task, the speed index metric for reaching toward a target task, and the CV of jerk and speed metrics for drawing a square and a circle tasks. In addition, BBT and a wide range of clinical and functional measures of activity and participation were performed (see Table 2). As some of the tasks included movements that were not trained, the described measures were also included as part of the transfer parameter (Table 2). Schaefer et al. (2013) measured the performance during a feeding task by measuring the number of successful repetitions, defined as spooning and transferring at least one bean from one cup to another. Transfer was assessed by sorting (BBT, spatiotemporally similar), dressing (spatiotemporally different), and dual tasks. The dual task conditions assessed whether automaticity transferred across tasks, measured by the difference between the reported and correct number of times a letter was heard in a sequence of letters.

Four studies examined the effect of different feedback frequencies on motor skill learning (Cirstea and Levin, 2007; Hemayattalab and Rostami, 2010; Hemayattalab et al., 2013; Burtner et al., 2014). Burtner et al. (2014) assessed the performance accuracy and consistency of an elbow extension-flexion reversal movement, which were assessed separately on acquisition at first day, retention (without feedback) and reacquisition (with feedback) on the second day. Accuracy was measured using the root mean square error (RMSE) (i.e., the average difference between the goal movement trajectory and the participants response), and consistency, by the variability of the RMSE. Hemayattalab and Rostami (2010) and Hemayattalab et al. (2013) evaluated new motor skill learning of children with CP—darts and bean-bags throwing tasks. The accuracy of scores was measured by the proximity to a center-peripheral target at acquisition and at retention, 3 days after practice. Cirstea and Levin (2007) examined the performance and retention following 1 month of pointing movements training to the contralateral workspace by measurement of the angular motions of the elbow and shoulder joints, the elbow-shoulder interjoint coordination and by the amount of trunk's anterior and rotational displacement. A transfer test included pointing movements toward an ipsilateral target.

Molier et al. (2011) examined the effect of position resistance feedback, provided when a deviation from a predefined path occurred, during three reaching task trainings (moving hand; making a curve; lifting hand to shelf). The average use of the feedback was calculated, and the difficultly level was established by measurement of the reached height and diameter of predefined path. In addition, elbow and shoulder joint excursions, positions, and coordination, and an isometric strength task for which the maximal voluntary torque (MVT) were measured during a task of circular arm movements. The parameters that were utilized to assess learning were not directly specified in the study. Therefore, as the circular movements were not trained, we placed the measures under the transfer principle (see Table 2). The measured feedback frequency and difficulty level during training corresponds with the performance change that also may be addressed as the acquisition parameter.

Kitago et al. (2013) aimed to determine the feasibility in implementing kinematic measurements (MT, PV, absolute initial directional error, path curvature, systematic error, number of submovements) of arm reaching and wrist pointing tasks, and of clinical measures for understanding the motor recovery process within 2 weeks after CIMT (i.e., retention. Chang et al. (2007) assessed the performance and retention of stroke subjects after reaching training, by measuring the PV, %TPV, MT, normalized jerk score, and limb muscle strength. Chen et al. (2014) assessed the performance and retention at 3- and 6-months by examining the speed and dexterity during 8 object manipulation tasks (Bruininks, 1978), by functional ability measures (see Table 2) and by kinematic analysis (RT, NMT, MS, PV, maximum grip aperture (MGA), and the percentage of movement where MGA occurs) during a reach-to-grasp task. Casadio and Sanguineti (2012) examined the performance change of stroke patients over a robot-assisted arm extension task practice. Performance measures (i.e., speed, precision, and smoothness), retention rate (dependence of voluntary control on previous trials), learning rate (dependence of voluntary control of next trials on current trial), assistance rate, noise (voluntary control not accounted for learning) and vision bias were measured. Retention was examined by estimating the correlation between the retention rate during performance to the percentage change in the FMA 3 months post the rehabilitation trial.

Five studies assessed motor sequence learning of stroke patients (Boyd and Winstein, 2001, 2004, 2006; Pohl et al., 2006; Orrell et al., 2007). Two of them included practice of the serial reaction time task (SRTT), which includes a movement to press one of four targets when cued. When the correct key is pressed, the next cue was delivered (Boyd and Winstein, 2001; Orrell et al., 2007). Boyd and Winstein (2001) assessed learning by performance change over practice of the median reaction time, whereas Orrell et al. (2007) assessed the median response time both during performance and retention. Orrell et al. (2007) also implemented two transfer tests, once by changing the motor sequence and by changing the required movement from index finger only to whole arm movement. Pohl et al. (2006) examined the performance change during practice of an implicit motor learning task of stroke patients undergoing a motor task with random and repeated sequences by the mean response time and by the CV of response time. To examine the performance following practice, subjects were requested to perform the sequence practiced in the repeated conditions once again. In one study (Boyd and Winstein, 2004) subjects practiced the continuous tracking (CT) task that included tracking of the vertical path of a target cursor. The middle third of each tracking trial was repeated, whereas the first and last third were random. Reduction in tracking errors measured by the RMSE and spatial-temporal accuracy were measured to assess performance change over practice and retention (Boyd and Winstein, 2004). In the final study subjects practiced both the SRTT and CT task (Boyd and Winstein, 2006). Learning was inferred by the median response time and RMSE respectively, at performance and retention.

Adaptation was assessed in the reviewed studies using dynamic perturbations by induced force fields during reaching movements (Dancause et al., 2002; Takahashi and Reinkensmeyer, 2003; Patton et al., 2006; Scheidt and Stoeckmann, 2007; Masia et al., 2011). Despite the variability in methodological designs (e.g., practice—number of trials and repetitions, utilized measures, etc.) between the studies, all induced a change in environment and assessed the resulted change in behavior. The manifestation of the adaptation parameter, in all studies, was evaluated by some sort of measurement of the extent and/or quickness in which the performance returned to the pre-perturbation level and of the sensory-prediction errors reduction. The changes in the after-effects assess the update of the internal model (Krakauer, 2006; Huang and Krakauer, 2009). Robots were described as suitable to record movement data (e.g., position, velocity, and joint torques), which allows quantitative reliable measurement of kinematics and dynamics during recovery (Huang and Krakauer, 2009).

The utilized measures and methods of their implementation can affect the obtained results. For example, Patton et al. (2006) concluded that stroke survivors preserved the ability to adapt, whereas Takahashi and Reinkensmeyer (2003) found their adaptive capability to be reduced. Their different results may be explained by the fact that Patton et al. (2006) utilized a metric that measured only the early part of the movement, whereas Takahashi and Reinkensmeyer (2003) assessed the whole movement. Evaluation of the anticipatory control should include a time limit for the movements performed. The limit's purpose is to minimize online corrections and focus on deficits in the feed-forward mechanism (Kitago et al., 2013). For example, Raghavan et al. (2006) estimated the predictive control of children with CP by estimation of the motor command at the first 70 ms of a grasp and lift task. Without the time limit, the prolonged time they needed to receive sensory feedback for grip force generation, would not have been identified (Raghavan et al., 2006). Contrarily, the findings of Patton et al. (2006) cannot serve as a measure of feed-forward control error as it does not consider the time between the feed-forward control and movement initiation (Patton et al., 2006).

It is interesting to note that only studies that explored the anticipatory control utilized measures of force generation, while other studies used only kinematic measures. This is true of all but four of the reviewed studies that did include force components as part of their measurements (Chang et al., 2007; Colombo et al., 2010; Bourke et al., 2015; Gilliaux et al., 2015). Takahashi and Reinkensmeyer (2003), who did not measure movement dynamics, suspected that the rate of force development explains the impaired anticipatory control of the paretic arms. Measurement of the force generation may clarify whether the deficit in the anticipatory control is due to inability to form internal models or to implement them (Takahashi and Reinkensmeyer, 2003).

Scheidt and Stoeckmann (2007) found that stroke patients adapt similarly to healthy individuals; however, they may require more practice, as they had more influence of prior error on subsequent movement. Dancause et al. (2002) added that severely affected individuals might require more practice, as they required more trials to diminish errors than mildly affected individuals. It can be inferred that measurement of the rate at which the errors decrease throughout trials and the influence of the prior error on the next movement, can be used as measures to evaluate the amount of practice an individual will require.

Adaptation may be an initial ingredient in a motor learning model (Bastian, 2008), and was implied to be a form of implicit learning that requires no awareness (Krebs et al., 2001). However, skill acquisition often requires conscious awareness and practice of the performance (Krakauer and Mazzoni, 2011; Kantak and Winstein, 2012). Performance curves indicating change throughout practice can be established for a variety of measures. Geerdink et al. (2013) used the performance curve of manual dexterity improvement to evaluate the point in time where maximal effects are learned and achieved, finding that age has an effect on the speed of dexterity gain. It can be inferred from these results that the learning curve can be used as a measure to study and evaluate the maximal effect of various interventions, which might later on assist to better establish individualized training timings to fulfill maximal potential.

None of the studies that examined the adaptation parameter assessed the persistence of the after effects. Measures of performance can be implemented at different times throughout the training, but can also be used at a time interval after the last practice to assess the retention. As performance might be affected by transient factors, such as feedback, attention, fatigue etc., performance measures may be limited to the acquisition phase. Therefore, a retention test is preferable to infer learning, as it assess long-lasting changes indicating constancy of the level of performance achieved at acquisition and strength of the motor memory (Schmidt and Lee, 2004; Kantak and Winstein, 2012).

Another important aspect of learning is the extent to which what was learned during practice generalizes outside of practice settings (Schmidt and Lee, 2004; Kantak and Winstein, 2012), assessed with transfer tests and reflects the flexibility of the motor memory (Schmidt and Lee, 2004; Kantak and Winstein, 2012). According to the motor learning theory of Fitts and Posner (1967), presenting a three stage model of learning—cognitive, associative, and autonomous, achieving transfer (automatization) and retention of a skill is necessary to indicate learning has occurred (Cano-de-la-Cuerda et al., 2015). It was suggested that when the trained and untrained tasks share similar neural demands, generalization to the untrained task is increased (Sainburg and Wang, 2002; Schmidt and Lee, 2004; Shadmehr, 2004). This implies that the transfer test should encompass and examine similar motor control requirements as the training. The transfer parameter is essential, since the rehabilitation aim is to grant the patient optimal function and independence, only possible when improvements generalize to real-life situations.

Some prominent theories of motor learning require a comprehensive analysis of movement that includes many of the measurements used in the reviewed studies (Cano-de-la-Cuerda et al., 2015). However, most of these studies did not qualify to infer learning according to these theories. For example, parameters and measures implemented during tasks with restriction on degrees of freedom contradict Bernstein's model of motor learning (Bernstein, 1967) that requires the learner to increase the degrees of freedom as an indication of learning. Similarly, many of the studies evaluated the efficiency and consistency of the movement without testing transfer and vice versa. Both are essential to demonstrate learning according to Gentile's theory of motor learning (Gentile, 1972). Therefore, studies who aim explicitly to demonstrate learning must confide to a theory, and take it into consideration when designing the study.

Improved activity can arise from either recovery of impairment, development of compensatory movements, or both. There is a necessity to differentiate between impairment of execution and impaired motor learning (Boyd and Winstein, 2004). The implemented settings and outcome measures of a study may affect the interpretation of the results. Patton et al. (2006) supported the hand against gravity, which could decrease the effect of the impairment in movement execution (due to weakness, spasticity etc.) and result in detection of adaptive capability. On the contrary, Takahashi and Reinkensmeyer (2003) did not support the hand, which might led to poor adaptive capability among subjects, when in fact execution impairment increased due to gravity.

The state of mind presented in a study by Nourrit-Lucas et al. (2013), who explored long term retention of neurologically intact subjects, might also be relevant to studies performed on neurologically impaired individuals. The authors suggested that the experimental settings in which the learning is assessed, affects the measurements and the conclusions drawn regarding the learning process. Performance variables usually assess simple tasks with few degrees of freedom. They are defined as parameters measuring performance in terms of speed and accuracy, representing the outcome of the behavior with respect to the goal of the task (Nourrit-Lucas et al., 2013). A more complex task has a larger number of degrees of freedom, and learning can be explored by coordination variables measuring the spatiotemporal functional organization between body segments in terms of phase relations (Kelso, 1997; Nourrit-Lucas et al., 2013). In most of the reviewed studies, simple tasks were examined, among the rest, to decrease the probability for compensation. Moreover, in approximately half of the discussed studies the arm was supported against gravity, reducing even more the degrees of freedom. Retention tests assess the same coordinative pattern: an improvement is expected due to the motor plan that was established during practice, whereas a transfer test would examine the adaptability of the motor program. Therefore, transfer tests are more suitable to examine the coordinative measures (Nourrit-Lucas et al., 2013). The coordinative improvement can represent the reacquisition of the pre-lesion patterns, addressed as recovery (Levin et al., 2009). None of the reviewed studies assessed the coordination of a more complex task.

Kinematic analysis of various sorts for measurement of parameters were performed by all but three of the reviewed studies (Hemayattalab and Rostami, 2010; Geerdink et al., 2013; Hemayattalab et al., 2013). Subramanian et al. (2010) separated the kinematic variables to measures of motor performance and to measures of movement quality. Two of the discussed performance measures that were often utilized in the reviewed studies to assess the parameters of motor skill learning are measures of accuracy and velocity of movement. Reduction in errors of these measures indicates improvement of performance (Krakauer, 2006; Schmidt and Wrisberg, 2008). However, in motor execution of a task there is a relationship between movement speed and accuracy, and depending on task requirements, the accuracy or speed component can be prioritized (Fitts, 1954; Kitago and Krakauer, 2013). If only one of them is assessed, an improvement does not necessarily indicate an improved skill. For example, a subject can make more errors as speed increases, or slow down for a more accurate movement (Kitago and Krakauer, 2013). The movement quality measures include the configuration of the examined limb and measurement of compensatory movements, and were suggested as able to distinguish between recovery and compensation (Subramanian et al., 2010). Almost half of the reviewed studies restricted compensatory movements, preventing examination of movement quality required in real-life settings. How then, can complex motor skills be assessed in terms of qualitative organization for neurologic patients with a variety of heterogenic impairments, potentially masking learning? In some of the reviewed studies the learning following a treatment was inferred by examination of the less affected limb (Table 1). This may distinguish motor execution impairments due to hemiparesis, which may mask motor learning, from deficient motor learning (Boyd and Winstein, 2004).

Many studies evaluate the performance ability of their subjects using functional tests (Kitago and Krakauer, 2013). Functional tests often include more complicated movements with multiple degrees of freedom, resemble real-life activities, and are categorized under the “activities” domain of the ICF classification (Sivan et al., 2011). However, clinical measures do not consider the quality of the movement, and therefore do not measure a decrease in impairment or return to a normal motor control (Kitago et al., 2013). Therefore, functional improvement in clinical measures, but not in measures of impairment (such as the FMA and kinematic analysis), can be attributed to compensatory strategies (Kitago et al., 2013). Kinematic variables were suggested to be valid for differentiating compensation from recovery and for measurement of upper limb impairment (Subramanian et al., 2010). Huang and Krakauer (2009) reviewed rehabilitation strategies and outcome measures for impairment versus function. They suggested that in the acute and sub-acute stages of recovery, rehabilitation treatment should focus on the impairment level while refraining from compensatory adjustments, and only after a certain level of improvement is achieved, focus on functional performance. It can be inferred that the measure of assessment should also be fitted to the stage and timing within the recovery process, and should be able to distinguish between them. Movement quality measures can be sensitive and useful for complementing clinical assessment (Subramanian et al., 2010). Moreover, Sivan et al. (2011) suggested that when evaluating a rehabilitation program it is important to measure each domain within the ICF classification. While most clinical tests were classified as evaluating the “activities” domain, most of the measures that we reviewed, such as kinematic analysis and the BBT, were placed under assessment of the body functions domain. This different classification might explain the absence or low correlation between the measures in some of the studies, as they might assess different features of learning. This implies that when assessing motor learning measures different features should be taken into consideration.

The validity and reliability are important components for a meaningful behavioral research. The validity represents the degree to which a study measures what it intends to measure, and reliability is the consistency of the results (Forzano and Gravetter, 2009). Therefore, both the validity and reliability are important in order to infer whether the parameters and measures used in the study provide accurate representation of the change during and after practice.

Kinematic movement quality measures were found to be valid and sensitive for recognizing upper limb impairments of stroke patients performing pointing and reach-to-grasp tasks (Subramanian et al., 2010). However, as mentioned previously, some of the reviewed studies did not examine quality measures, due to restriction of the upper limb during movement and of compensatory movements, or due to measurement of solely motor performance measures.

While the validity and reliability of the clinical measures were mostly stated in the reviewed articles, these were less addressed for the motor learning parameters and laboratory-based measures. Some studies mentioned the objective behind their motor control outcome measures. For example, Gilliaux et al. (2015) and Durham et al. (2014) measured kinematics previously established and described sensitive, respectively Geerdink et al. (2013) used the BBT because of its feasibility, validity and reliability (Jongbloed-Pereboom et al., 2013). Other studies focused on examining the strength of their measures. Bourke et al. (2015) found good to excellent reliability coefficients for task performance measures. Kinematic analysis was described as sensitive for motor control measurement (Chen et al., 2012), and for evaluation of motor recovery (Caimmi et al., 2008). Kitago et al. (2015) suggested that their method of analyzing reaching kinematics, based on functional principal component analysis (Yao et al., 2005; Goldsmith et al., 2013), is more sensitive than measures such as the end-point accuracy or PV kinematic measures due to its additional ability to evaluate the entire trajectory of movement.

Across the reviewed studies, not all parameters were equally studied and the reasoning for parameter selection was not addressed in all but one study. The validity and reliability of the assessed parameters were also not addressed. Kantak and Winstein (2012) suggested that retention and transfer tests, rather than solely the performance parameter, should be implemented to negate transient performance changes and detect motor learning. In this review, the performance was most often assessed, in 24 of the reviewed studies, whereas transfer was evaluated in 14, and retention in only eight of the studies. Retention was examined following a time interval in all studies, but the length of the retention interval varied. Kantak and Winstein (2012) categorized retention into immediate and delayed (i.e., implemented at least 24 h after practice), and suggested that different retention intervals can yield different conclusions regarding the obtained learning (Kantak and Winstein, 2012).

This review has several limitations. Firstly, we focused only on studies of the upper limb. As the lower limb role and function differs from the upper limb, it is important to further address the metrics utilized for assessment of the lower extremity with relation to their place within the motor learning model. Secondly, only studies assessing CP and stroke subjects were included. Important information regarding the parameters and underlying measures might have been already studied in other populations and settings that were excluded from the review. It is also possible that we missed additional relevant studies when searching the databases, despite the multiple key words, due to inconsistency in terminology. Furthermore, the included studies were very heterogenic in their methodology, motor tasks, and measures. We categorized the identified measures according to the evaluated parameters. We could not synthesize the findings across the studies to a better extent due to the measures' high variability between the studies. Due to the variability in study design, and the narrative nature of this review we did not assess the methodological quality of the studies. Despite the broad utilization of kinematic analysis in the reviewed studies and despite reports of validity and reliability of some measures, we could not evaluate the appropriateness of measures' selection and the quality of measurements. An additional drawback is that we examined the evidence for the reliability and validity only within the reviewed studies, possibly missing other studies whose sole purpose was to establish these features rather than examine the outcome of a treatment. We suggest that a review with that scope in mind should include additional search of the databases with focus on the reliability, validity, sensitivity, and specificity of the implemented measures, while taking into consideration the variety in the characteristics of studies. This may serve to establish an algorithm for the selection of outcome measures.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The reviewer PL and handling Editor declared their shared affiliation, and the handling Editor states that the process nevertheless met the standards of a fair and objective review.