Principles of Neurorehabilitation After Stroke Based on Motor Learning and Brain Plasticity Mechanisms

1So far there is no clear understanding of the principles underlying effective neurorehabilitation approaches. Therapeutic protocols can be readily described by the following aspects: the body part trained (e.g., the legs), the tools or machines used for the training (e.g., a treadmill), the activity performed (e.g., walking), and when the therapy commences (e.g., during the acute phase after a stroke). However, an intervention typically includes more elements. For instance, the use of the less affected limb can be restricted, and the therapist can encourage the patient to spend more time exercising or give feedback about task performance. While some interventions, like CIMT, clearly define their active ingredients that should lead to effective recovery, most others do not. Neurorehabilitation research aims to find interventions that promote recovery and to establish whether the presence or absence of improvement can be explained by any neuronal changes that occur in the post-stroke brain. Neuroscience can help us to create interventions that lead to changes in the brain; however, with no clear understanding of what an intervention does, attributing causality remains difficult. One way to formalize an intervention is by breaking it into parts, studying the behavioral and neural effects of these parts, and deriving principles from them–in the case of stroke neurorehabilitation, these would be principles that optimize acquisition, retention, and generalization of skills.

2While there are plenty of meta-analyses that look at training effectiveness in terms of individual body parts/functions, tools, or machines and activities, the effect of experience remains much less clear in spite of attempts to formalize and identify the principles of neurorehabilitation. A review of principles of experience-dependent neural plasticity by explains why training is crucial for recovery. According to their work, neurorehabilitation presumes that exposure to specific training experiences leads to improvement of impairment by activating neural plasticity mechanisms. Consequently most of the work in the field focuses on the identification of scientifically grounded principles that should guide the design of these training experiences. In this vein, elaborated on five main principles of effective training experience — specificity, repetition, intensity, time, and salience — but offered little concrete applicability. Another synthesis addressed further principles (forced use, massed practice, spaced practice, task-oriented functional training, randomized training); however, the main focus of the review was on individual body functions, methods, or tools, providing a global view on rehabilitation strategies. Two meta-analyses investigated specific principles. One looked only at the principle of intensity and found that more therapy time did enhance functional recovery. Another determined that repetition does improve upper and lower limb function. However, both studies did not investigate the mechanisms that would lead to the effects observed. Similarly, a review that analyzed CIMT, which combines several principles in one method, gained interesting insights in its efficacy but did not explain the results from a neuroscientific, mechanistic point of view. The work by, on the other hand, tried to link the principles of motor learning to the application of these principles in novel rehabilitation methods while offering some neuroscientific reasoning for doing so. Their review addresses the difficulty of the task, the organization of movement, movements to the contralateral workspace, visual cues and objects and the interaction with them, sensory feedback, feedback about performance and results, repetitions, variability, and motivation. However, the included motor control and motor learning principles were not well defined and therefore leave room for interpretation.

3In a previous meta-analysis, we compiled a list of principles for neurorehabilitation based on literature on motor learning and recovery: massed practice, dosage, structured practice, task-specific practice, variable practice, multisensory stimulation, increasing difficulty, explicit feedback/knowledge of results, implicit feedback/knowledge of performance, movement representation, and promotion of the use of the affected limb. We then performed a content analysis to determine whether these principles were present in the clinical studies included in the review, but we did not provide an analysis of the principles identified. In this work, we aim to extend the number of principles found and, for each of them, unify the neuroscientific literature from human or animal studies on motor learning and comment on the observed neuronal effects. We also include evidence from clinical studies to show its effect in recovering functionality after stroke. Some principles already serve as building blocks of effective rehabilitation programs, e.g., CIMT, Bobath, enriched rehabilitation, VR-based rehabilitation, and exogenous or robotic interventions. However, transferring these principles into clinical practice faces the challenge of operationalizing them. We comment on these difficulties and the gaps between theory, evidence, and operationalization that we encountered. Consequently, this work can serve clinicians and researchers as a practical guide of principles to investigate further effective neurorehabilitation approaches.

4In this conceptual analysis, the rehabilitation experience is broken down into individual parts that are termed principles of neurorehabilitation. They are principles because they are evidenced by experimental data, and together, they could form the foundation of a higher-order theoretical framework. As a first attempt, a list of 11 principles was compiled based on existing literature in a recent meta-analysis on the effectiveness of VR-based rehabilitation systems for stroke recovery. For the current work, the list has been revised, and additional principles have been identified through a computerized search in PubMed Central using the keywords “principles of motor learning,” “principles of recovery,” “principles of experience-dependent learning” and “principles of neurorehabilitation.” We restricted the search to the last 5 years to obtain currently used principles. We focused on reviews, perspectives, and debates around rehabilitation methods and interventions for stroke recovery and excluded articles that explained study protocols or clinical trials, prevention methods, pharmaceutical or medical interventions, or stroke taxonomies. The principles mentioned in each paper were compared with the original list and added if they were not present. Afterward, we summarized for each principle the historical background based on motor learning literature and its contribution to learning based on human or animal studies. Further, where available, neurological effects and clinical outcomes were included as well.

5Our computerized search yielded 548 records, of which 74 were deemed adequate for further screening after we examined if their titles either contained any of the search terms or appeared to discuss post-stroke rehabilitation strategies. After analysis of their abstracts and full-texts, the principles mentioned in 17 articles were extracted. We excluded papers if their title or abstract reported or compared surgical or pharmaceutical interventions as well as if they discussed stroke taxonomies, proposed study protocols or clinical trials, covered principles unrelated to stroke and/or stroke rehabilitation itself (e.g., principles for disease prevention, pre- and post-operative care, care facilities, patient management, therapist education, nursing practice, dietary recommendation, veterinary etc.), or looked into patient or caregiver perception. The articles and reviews selected spawned various research fields in neurorehabilitation: Motor learning, therapies [physical therapy, upper limb immobilization, environmental enrichment, aerobic training, CIMT, cognitive rehabilitation, music therapy], tools and methods [hand robotics, VR, neurofeedback], and principles [dose and timing]. Together with previously collated literature, we identified 15 principles.

6The identified principles from the meta-analysis are as follows:

7Massed practice/repetitive practice.

8Spaced practice.

9Dosage/duration.

10Task-specific practice.

11Variable practice.

12Increasing difficulty.

13Multisensory stimulation.

14Explicit feedback/knowledge of results.

15Implicit feedback/knowledge of performance.

16Modulate effector selection.

17Action observation/embodied practice.

18Additional principles encountered through the search:

19Goal-oriented practice.

20Rhythmic cueing.

21Motor imagery/mental practice.

22Social interaction.

23In the following sections, we summarize for each principle the theoretical background, the evidence for motor learning, and the clinical effectiveness. We also added studies that comment on the neurological changes observed after applying the principles in motor learning tasks. The detailed neurological changes reported by these studies can be found in Table 1.

24Massed practice was defined as work episodes with very brief to no rest periods. Within a work episode, a skill can be trained repeatedly in a constant or blocked fashion. In the field of rehabilitation, the term describes the prolonged and repeated use of the more affected limb. Theoretically, learning through repetitions can speed-up the shaping of priors, which, together with likelihoods based on sensory input, aid in making an optimal estimate for action selection. Animal studies have shown that repeating skilled movements leads to localized changes in the area responsible for the movement, whereas the pure repetition of unskilled movement does not. In humans, early studies have shown that blocked practice leads to faster acquisition, but poorer retention and less transfer than variable practice and that massed practice without breaks seems less effective for motor performance.

25In standard therapies or clinical studies, the amount of repetition is typically not quantified but was observed to be an order of magnitude lower than in studies investigating recovery in rats and monkeys. Instead, the evidence for massed practice relies typically on the number of sessions or duration. A study looking into the feasibility of translating repetition amounts of animals to humans found improved motor functioning after training with high-repetition doses. However, no “pure” repetition training was provided, as the protocol included a variety of tasks that increased in difficulty. On the contrary, a study comparing four groups with different repetition amounts did not find significant differences based on the number of repetitions. This intervention included other principles as well. Meta-analyses confirm the mixed effects of repetitive training on improvement. Hence, massed practice appears to be a commonly used ingredient, but its clinical operationalization is often confounded with other principles. In order to investigate its true effects on recovery and compare across studies, the repetitions within a training session and across therapy duration should be measured and quantified.

26Spaced practice implies that training should be structured in time to include rest periods between repetitions or sessions. Instead of spaced practice, the term distributed practice is often used in literature. However, some authors use the term distributed practice as a combination of spaced and massed practice. Research on human skill acquisition suggests that increasing the time spacing between learning periods improves final test performance. However, when these learning periods are too long, learning and retention rates drop. The mechanisms behind the effects of distributed practice remain unclear. It has been hypothesized that the first exposure to a stimulus pre-activates its representation in memory, requiring no further activation in a subsequent repetition trial, leading to a poorer internal representation of that stimulus, which has been termed as the repetition suppression effect. Animal and fMRI studies support this hypothesis, showing that neuronal activation decreases after stimulus repetition where the magnitude is modulated by the delay between the first and second presentation, with larger delays leading to greater decreases. Spaced practice might counteract the repetition suppression effect by canceling stimulus priming. TMS revealed that primary and supplementary motor areas are involved in motor memory consolidation, which might be facilitated by spaced practice. Further, learning and physical activity have been linked to hippocampal neurogenesis. Animal studies also suggest that spaced practice facilitates long-term memory formation by fostering the survival of cells in the dentate gyrus that are important for learning and memory. Also, in vivo spacing of electrical stimulation facilitates the recruitment of protein-synthesis-dependent processes, which facilitates late LTP effects.

27In the clinical field, only a few studies have investigated the effect of spacing on post-stroke recovery. A clinical study that investigated whether a CIMT protocol could be distributed over more days with less therapy time per day showed improvement in motor outcomes that were similar to previous CIMT protocols and superior outcomes in long-term quality of life.

28Unlike in pharmacology, dosage is an ill-defined term in rehabilitation. Generally, it is operationalized as the number of hours spent in therapy, the frequency of training sessions and the duration of a session, or the training amount required to stimulate learning. High dosages are often equated with high intensity of training. However, the intensity of training could also be operationalized as the metabolic cost, work rate, or perceived intensity through exertion, which are rarely measured in standard therapies except in fitness and aerobic protocols.

29Typically, inpatients receive only 22 to 60 min of training a day, with fewer minutes at later stages. There is some evidence that increasing therapy hours would be beneficial to speeding up functional recovery. At least 16 h of extra training (e.g., 71 more minutes per day for 3 months) within the first 6 months seem to be required for functional gains. However, there is some controversy over the benefits of increased training early after stroke, and a pooled analysis revealed no evidence of an effect of additional doses. Hence, the exact dose-response for different therapies at different stages post-stroke needs to be determined. Also, it seems that motor performance needs to reach an asymptotic level in the first session to facilitate delayed performance gains across sessions or days. Therefore, delayed performance gains seem not to depend on repetition or over-night consolidation, but on the amount of training that induces asymptote in the individual’s performance. Neurologically, high-dose rehabilitation protocols with extended training hours possibly induce structural plastic changes as well as a reorganization of neural networks (summarized by ), increase cortical excitability and improve motor function and use. Several studies observed a normalization in ipsilesional cortex activity, which could underlie the functional gains.

30Task-specific practice postulates that changing the conditions of a task might require a change in the abilities needed to execute it; conditions during training should match the conditions during testing. Thus, the specific conditions of practice shape the internal sensorimotor representation of the skill learned, leading potentially to highly specialized skills whose performance is superior in transfer tasks that meet the training conditions. Grounded in this principle, conventional rehabilitation protocols focus their training on the execution of ADL, as they are deemed meaningful to the patient. Since the main target of rehabilitation is to enable the patient to perform ADL independently, therapy might not prioritize the restoration of pre-stroke movement patterns but allows the patient to acquire compensatory movement skills.

31One study with a large sample size found that task-specific practice appears to be similar to standard therapy in improving motor functionality. On the other hand, smaller fMRI studies found that task-specific training facilitated motor learning and retention and induced a change in the laterality index, which was confirmed in other studies as well. However, while two studies found reduced activity in the contralesional cortex, one found changes in neuronal activity patterns in both hemispheres. A study with TMS demonstrated a trend toward reduced interhemispheric inhibition following task-specific training.

32Since a given goal (e.g., throwing a ball into the basket) could be accomplished by many different motor synergies, it is assumed that movement control is achieved through the coupling of goal-specific functional movements. Goal-oriented practice, therefore, does not emphasize primarily individual muscles or movement patterns involved in execution but requires the patient to explore the couplings that are suitable to achieve the task. In general, motor skill performance and learning are enhanced if attention is directed to the effect of movement instead to the movement itself. Goal-oriented movements appear to produce a better reaching performance than the same movements without a goal, and setting specific, difficult goals leads to higher motor learning performance than non-specific goals. It appears that probing a skill in a goal-directed fashion after overnight consolidation promotes better performance than probing the skill by drawing attention to finger movements. Evidence from studies looking into tool-use in animals and humans suggest that, neurologically, action goals are represented as effector-dependent in the anterior intraparietal sulcus and primary motor areas, and as effector-independent in the ventral intraparietal sulcus and premotor cortex. Goal-oriented movements produce higher activity in sensorimotor areas.

33There is some evidence that goal-oriented practice is beneficial for recovery. However, the described interventions seem to be confounded by other principles that are sometimes ascribed to goal-oriented training.

34Variable practice can be achieved in two ways: (1) by providing variability within a training sequence, a method termed as variability of practice, or (2) by randomizing the presentation of individual training sequences, a method termed as random practice or contextual interference. Both methods have been shown to lead to better retention and enhanced generalization to similar but untrained tasks or movements, despite hampering initial performance. However, a random presentation of information might be detrimental to motor learning. Imaging studies have shed some light on the mechanisms supporting these effects. fMRI and TMS studies in humans indicate that improved performance due to variable practice correlates with increased neuronal activity and connectivity in the areas of the motor learning network during acquisition, which is associated with better performance at retention stages. Also, the motor cortex showed greater excitability during retention. These results point to more efficient retrieval of motor memory due to variable practice. More complex bimanual visuomotor tasks that were practiced randomly have shown modality-specific activation patterns that led to the recruitment of areas related to visual processing. The effect of variable practice might be related to the strong link between the neuromodulatory systems that control neuronal plasticity and novelty, for instance, the dopaminergic, cholinergic, and noradrenergic systems, which are used by the brainstem activation system for controlling the global state of arousal.

35In the clinical context, one study that investigated random versus blocked practice failed to find an effect. It seems that this principle is rarely studied explicitly in clinical studies, but instead applied in conjunction with other principles to overcome boredom.

36According to and based on the ideas from, task difficulty can be described by the training requirements and conditions that are pertinent to the task, called the nominal task difficulty, and by how challenging the training is relative to the skill of the performer, called the functional task difficulty. Practice leads to fewer prediction errors and less need to process error information. Increasing the nominal task difficulty hence increases prediction errors and error processing demands. The optimal challenge point lies where functional task difficulty leads to a balance between information processing demands and performance, which is optimal for learning. It has been shown that training with difficulty levels personalized to the learner’s capabilities leads to superior learning outcomes than when increases in difficulty are fixed. Further, if subjects can control the task difficulty by themselves, their motor performance during acquisition and retention is significantly better. However, if difficulty surpasses one’s perceived ability to succeed, it might lead to detrimental effects on performance. Brain imaging studies showed increased activity in lateralized pre-motor and sensorimotor areas, but with an even more pronounced increase in parietal areas, pointing to a specialization of that area for task complexity. Potentially, noradrenergic neurons keep track of high or low task performance due to difficulty by switching their activity pattern preceding behavior.

37In stroke rehabilitation, task difficulty has been partly investigated through shaping or graded practice. Shaping is a concept that was initially used by behaviorists studying operant conditioning in animals and that was successfully transferred from animals to humans by making it part of CIMT: The use of the impaired limb is augmented by progressively increasing the complexity of the required movement. Although shaping appears to be one of the essential components of CIMT, its particular effect on motor recovery has not been studied on its own. Increasing difficulty has been successfully used in standard care studies, robot-assisted therapy, and VR-based systems, all of which showed beneficial effects on motor recovery. Task difficulty appears to be implicitly present in many tasks that investigate motor learning without being explicitly operationalized.

38The perception and integration of multiple senses are fundamental abilities of the brain. Because sensory information is noisy, the integration of various modalities requires probabilistic estimations to enhance perception. Studies in the cat superior colliculus showed that a single neuron could be responsive to several sensory modalities. In primates, the classic areas associated with multisensory processing are the superior temporal sulcus, the intraparietal cortex, and the frontal cortex, with newer studies confirming multisensory processing also in areas that were previously thought to be mainly unisensory. One sensory input (e.g., touch) can influence how another sensory modality is perceived (e.g., vision); therefore, exposure to multisensory feedback can enhance the ability to detect, discriminate and recognize sensory information. For instance, active physical exploration of multisensory stimuli led to greater accuracy in an associative recognition task showing enhanced connectivity between sensory and motor cortices. Animal studies demonstrated that sensory feedback is crucial in motor learning. Monkeys with an ablated primary sensory hand area had no problems in executing a previously known task but were unable to learn new skills. Providing multisensory stimulation during goal-oriented action execution might help to establish sensorimotor contingencies. Muscle vibrations appear to influence the sensorimotor organization, whereas paired associative stimulation with TMS increases motor-evoked-potentials.

39Of specific interest for rehabilitation is the integration of visual and proprioceptive information to perform movements. It has been shown that vision and proprioception are weighted differently at various stages during motor planning, suggesting a target for multisensory manipulations. Concurrent haptic feedback during motor imagery appears to enhance the classification accuracy of brain-computer interfaces when decoding movement intention, indicating that it can aid in closing the sensorimotor loop. Multisensory stimulation training might help patients to recover from unimodal deficits, for instance, visual deficits or auditory localization deficits.

40Neuroentrainment encompasses the study of the temporal relationship between the body’s movements and the rhythmic stimulation emerging from the environment. Any sensory modality (auditory, visual, tactile, or vestibular) can be used for entrainment. To date, there is not much literature about visual entrainment, possibly because the auditory-motor synchronization appears to be mainly driving internal rhythmic movement control. Hence, mainly auditory cues are used to synchronize movements to rhythmic patterns. Rhythmic patterns act like a template whose sequence can be anticipated. The regularity detection and tempo tracking of rhythmic patterns increases the activity in motor network areas and cerebellum and creates a mental representation of the rhythm, the so-called auditory model, which enables motor movements to anticipate the rhythmic pattern. The pooled evidence provided in the reviews by and suggests that there are neuronal interactions between auditory and motor systems, and auditory-cued motor training can change their mutual structural connectivity. The auditory-motor action coupling relies on a subcortico-thalamic-cortical circuitry that can be activated through extrinsic cueing. Cerebellar patients cannot consciously perceive rhythm changes and show high variable motor responses. However, rhythmic synchronization, respectively, motor entrainment remains intact, suggesting that the cerebellum might control the rhythmic auditory-motor synchronization by monitoring rhythmic patterns. Even without cueing, repetitive movements become periodic over time, as observed when analyzing gait patterns. The gait impairment observed in PD is ascribed to a deficiency of the internal timing ability that disturbs coordinated rhythmic locomotion, and which can be improved with rhythmical auditory stimulation. Besides, rhythmic somatosensory cueing of stride frequency through vibrotactile stimulation at the wrist could improve qualitative walking performance in PD.

41There is evidence that auditorily paced treadmill walking can improve gait coordination in stroke patients as well. Further, bilateral arm training with rhythmic auditory cueing enhances functional motor performance, which is maintained long-term and induces cortical and cerebellar changes. Meta-analyses found large effects that rhythmic auditory cueing improves walking velocity, cadence, and stride length and beneficial effects on improving upper limb impairment and function after stroke.

42KR has been defined as verbal, terminal and augmented feedback about goal achievement. Although the finding that extrinsic feedback can effectively create simple stimulus-response associations was brought forward by animal research in reinforcement learning, KR signifies more than just extrinsic rewards. KR contributes to learning through cognitive processing, not through conditioning. KR is provided through explicit feedback. Explicit feedback is given on quantitative or qualitative task outcomes, e.g., correctness, exactness, success, or failure. This feedback does not have to be verbal. For instance, when failing to reach for a target, the subject can hear unpleasant tones or see that the failed targets change color. Also, explicit feedback about kinematic outcomes can be KR, e.g., playing back a recorded movement after execution. However, this feedback supports learning only if the movement features that led to the outcome are pointed out to the subject. Explicit feedback seems to activate explicit learning mechanisms and shows only subtle effects on implicit learning mechanisms. While implicit learning appears to increase the cortical motor output maps of the involved movement initially, they return to baseline topography once the learned content can be explicitly declared. Possibly through explicit feedback a global motor plan is learned that is represented by higher-order neuronal networks, which influence the cortical sensorimotor representations differently. Rewarding or punishing feedbacks appear to have dissociative effects on skilled motor learning. Punishment can speed up motor learning, whereas rewards ensure long-term retention. The reinforcement of positive outcomes appears to foster a success-driven learning system, which limits decay after learning, possibly by mobilizing the dopaminergic system. Reward expectations modulate the activity of caudate neurons (striatal projection neurons), which receive reward-related information through the dopaminergic input from substantia nigra and spatial information through the cortico-striatal connection. Consequently, they modulate the inhibitory output of the basal ganglia, biasing attention to rewarded items. Either reward-driven activity of caudate neurons is a result of cerebral plasticity, or activity in the cerebral cortex is influenced by caudate neurons through the output nuclei of basal ganglia. Dopamine has a gradual build-up and can persist for longer time courses; it might support long-term memory formation of motor actions.

43KR has been used to reinforce adherence to CIMT. Meta-analyses often analyze KR together with KP under the umbrella term augmented feedback. A meta-analysis analyzing different feedback types reported positive effects on motor function for KR. However, this evidence is based on one study, whose intervention included other principles as well. It can, therefore, not be established whether KR is effective for motor recovery.

44KP was defined as feedback given about movement execution in the form of verbal descriptions, demonstrations, or replays of recordings. Advances in technology made it possible that KP can be delivered online, in an implicit manner and concurrent during movement execution, providing verbal or non-verbal feedback about ongoing intrinsic somatic processes and movement kinematics. For instance, feedback in the form of sounds and colors can be given while trunk displacements surpass a threshold. Biofeedback uses physiological sources like electromyograms to provide patients with real-time visual or auditory signals about their motor activity. Ultimately arm movements can be visualized and augmented using VR representations.

45Implicit sensory feedback enhances learning from sensorimotor prediction errors, which for instance can aid the adaptation to unexpected perturbations, possibly by contributing to implicit learning mechanisms. Concurrent implicit feedback leads to lasting adaptations to visuomotor rotations, which are not or less observed when feedback about movement outcome, e.g., KR is given. Although KP appears to be beneficial during training, there is evidence that subjects can become dependent on it, showing inferior performance when feedback is removed. compared the effects of providing concurrent visual to concurrent auditory feedback during the acquisition of a bimanual movement pattern. The authors found that subjects that had obtained visual KP showed poorer performance during retention testing than subjects that were given auditory KP. During acquisition, the visual feedback increased the activity in vision/sensorimotor-specific areas, which was maintained during retention testing even in the absence of feedback. On the contrary, the concurrent auditory feedback reduced the activity in temporo-parieto-frontal areas and deactivated task-specific sensory areas during retention testing without feedback. These results suggest that subjects can become dependent on concurrent visual feedback, but not on concurrent auditory feedback because they rely on sensory processing areas that have become tuned to visual information during practice. The auditory feedback, on the other hand, might foster the formation of an internal controller, evidenced by the stronger activation of prefrontal areas. Alternatively, auditory feedback might promote reliance on proprioception and is consequently ignored during training. Results from cerebellar patients that were exposed to force-field learning tasks propose that the cerebellum may play an important role in using implicit information to correct and adapt motor commands to changed limb dynamics, and in forming internal controllers. In contrast to explicit error signals mediated through midbrain dopamine neurons in basal ganglia, implicit sensorimotor errors are possibly encoded by cerebellar climbing fibers and manifest in complex spikes in Purkinje cells during reaching tasks. Computational modeling of adaptation to visuomotor rotations following concurrent visual feedback points to narrowly tuned neurons in the cortex that are driven by a prediction error that is computed by the cerebellum.

46Stroke patients experienced a significant recovery in motor function and showed increased activation in the ipsilesional primary sensorimotor cortex after 4 weeks of training with a VR system that provided them with implicit feedback about their upper-limb movement. However, the system also included several other principles. In addition, the provision of KP has been shown to recover impaired movement patterns, to reduce learned non-use, and to lead to longer-lasting recovery effects. A meta-analysis found a beneficial effect for KP on motor function; however, the effect was based on two studies only.

47In the acute stage after stroke, patients typically suppress the use of the affected limb due to pain, weakness, or malfunctioning. As a consequence, they are prone to overuse the non-paretic limb, and the resulting under-usage of the impaired limb can cause a loss of behavioral and neuronal function. Some authors argue that this compensation strategy, called learned non-use, emerges because the spontaneous use of the paretic limb does not cross a threshold level. Although standard therapy focuses on improving the functionality of the impaired limb, the improvement does not transfer to increased use of the arm for ADLs.

48Of those therapeutic approaches that were successful in counteracting learned non-use CIMT is the most common and most successful one. An fMRI study revealed changes in brain activity patterns due to paretic arm use in patients that underwent a 2 weeks CIMT program at home where the non-affected arm was constrained for 90% of the waking time. Increased grip strength in the affected limb correlated significantly with increased fMRI signal change in ipsilesional cortico-cerebellar areas. However, a meta-analysis did not find a pooled effect that forcing the use of the paretic arm alone is effective. Other approaches aimed at promoting paretic arm-use through positive reinforcement during bilateral arm training or through wearable devices (e.g., bracelets) that provide feedback about performance of ADLs.

49Action observation gained increased attention after the discovery of mirror neurons: in monkeys, some neurons discharged not only when the animal executed a motor command but also when it observed another individual executing it. In humans, subjects who first observed other individuals performing a novel task performed better in the same task than control subjects that did not observe other individuals or observed a slightly different task. It is thought that in monkeys, as in humans, action observation relies on the frontoparietal network. Indeed, a meta-analysis showed that in humans, movement observation, as well as movement execution, recruits mainly the premotor and parietal areas. Movement observation, however, exclusively activated the visual cortex, whereas execution activated the primary motor cortex. Therefore, action observation might facilitate movement execution and motor learning by facilitating the excitability of the motor system. Indeed, TMS during action observation elicited increased muscle activation patterns. For practical reasons, action observation could be especially beneficial for stroke patients with severe hemiparesis or complete paralysis. There is some clinical evidence that action observation therapy can reduce impairment and increase brain activation in the frontoparietal network and bilateral cerebellum.

50Besides internalizing someone else’s movement, humans can also ascribe ownership and agency to body parts that do not pertain to them. The discovery of rubber hand illusions led to insights about the mechanisms underlying agency. Both the sense of agency and ownership are susceptible to manipulations, that have been used for therapeutic purposes, for instance, in mirror therapy. Similar to action observation, mirror therapy appears to rely on the frontoparietal circuit, which is why its motor learning effects are partly explained by the same mechanisms. However, contrary to movement observation, mirror therapy robustly activates the primary motor cortex and visual processing areas ipsilateral to the mirrored movement. Also, mirror therapy seems to increase functional connectivity between cortical motor areas and to excite the neural connection between the two hemispheres. A meta-analysis attests mirror therapy a significant long-term effect on motor function, the ADLs, the reduction of pain and the reduction of visuospatial neglect.

51If the impairment of the limb impedes active movement, visual illusions could be presented to the patients to simulate movements with the paretic arm. The error-prediction mechanism driven by the cerebellum could be equally activated through the alternative representation. Possibly, the stronger the visual illusion, the more agency is ascribed to it, which could explain the difference in brain activation patterns between action observation and mirror therapy. The sense of agency seems to be important when learning from sensorimotor prediction errors, respectively agency is reduced when prediction and outcome do not match. However, there is no consensus on the definition of ownership and agency, which makes their operationalization in clinical practice difficult.

52Mental practice and motor imagery rely on the ability to simulate actions mentally without overt behavior, as summarized by the simulation theory. Motor imagery can be seen as a mental rehearsal of future movements and motor plans, that can be beneficial for motor learning. However, actual physical practice shows superior effects on learning. A meta-analysis compared the brain areas that are active during mental imagery and movement execution. Both seem to recruit premotor areas, somatosensory cortex, and subcortical areas. Also, activation in the mid-cingulate cortex was found, with motor imagery activating more the anterior region that is linked to the cognitive aspects of motor control, whereas motor execution recruiting more the posterior region that is associated with basic motor functions. While motor imagery appears to activate more the parietal cortex, movement execution appears to recruit more classic sensorimotor regions like the primary motor cortex and cingulate motor areas. These findings are in line with studies showing that lesions in the frontoparietal system can diminish the ability of motor imagery. Motor imagery and physical practice also appear to induce similar learning-dependent brain changes. Not surprisingly, the activation pattern of motor imagery appears to be similar to the one identified in action observation and mirror therapy.

53The learning effects of motor imagery and mental practice have been extensively studied in sports, whereas research regarding their clinical efficacy and efficiency is sparse and relatively recent. However, motor imagery is thought to be advantageous for stroke recovery, especially for severely impaired patients. Since patients retain the ability to imagine movements with the paretic limb, mental motor practice might facilitate functional reorganization. A meta-analysis looking into the effectiveness of mental practice also found some trends for positive outcomes. However, pooled effects could not be estimated because only a few Class I studies exist, and their protocols, measurements, and interventions vary widely.

54Social interaction has been defined as a behavior in which the participants’ actions are both a response to and a stimulus for the counterpart’s behavior. Many ADL implicate social interaction, and a failure to perform them might lead to an undesired dependence on others. The level of self-efficacy influences motor skill performance and learning, and in turn, is influenced by the appraisal or discouragement from others. fMRI recordings of a subject experiencing a live social interaction revealed activations in areas commonly identified in the perception of social cues besides other regions involved in goal-directed and visual attention as well as reward processing.

55Animals that are allowed social interaction when recovering from an artery occlusion show higher functional improvement, increased recovery of behavior, and lower mortality, especially if the interaction partner was healthy. Including and investigating the impact of social interaction as part of the rehabilitation experience seems an important but missed opportunity. We found no study that was evaluating this specific aspect in a randomized controlled trial. One study evaluating enriched environments that included social interaction found positive results in terms of activity.

56This synthesis aimed at identifying a set of principles that should guide the design of effective neurorehabilitation protocols for post-stroke recovery. We identified 15 principles based on existing work on motor learning and recovery: massed practice/repetitive practice, spaced practice, dosage/duration, task-specific practice, task-oriented practice, variable practice, increasing difficulty, multisensory information, rhythmic cueing, explicit feedback/knowledge of results, implicit feedback/knowledge of performance, modulate effector selection, action observation/embodied practice, mental practice, and social interaction. Where possible, we identified the therapeutic and neurological effects of these principles from experimental work and clinical studies and commented on their limitations.

57Our motivation for this analysis is twofold. Firstly, we are confident that the quality of evidence from clinical work and its interpretation would be enhanced if interventions are described along with the included principles. Reviews or meta-analyses with ambiguous effects often state that the included protocols remained vague on the exact experience provided to the patients, which makes the comparison and interpretation difficult. By focusing solely on the ingredients of therapeutic interventions and compiling their current neuroscientific evidence, we aim to raise awareness of their importance. Also, this work might serve as a guide for clinicians and researchers to construct or identify the active ingredients in their interventions and to discover evidence currently missing. Secondly, we believe that there is a need to create a link between the principles of motor learning and their current operationalization in clinical studies and practice. We have identified several difficulties and shortcomings that do not aid in obtaining a common understanding of these principles and hence complicate the clinical investigation.

58It seems that many principles are poorly operationalized in clinical trials. For instance, when massed practice is investigated, the repetitions performed within a session and during the treatment duration are rarely quantified such that recovery effects due exclusively to repetition cannot be singled out. Also, the clinical research of spaced practice and dosage/duration would benefit if the parameters were quantified in a standardized way. Particularly dosage should be explicitly described in treatment minutes per session in order to be able to establish a dose-response due to training. Furthermore, dosage/duration should not be equated with intensity since the intensiveness of training cannot be estimated through treatment minutes only. Intensity should be an independent principle that needs to be investigated separately. Task-specific and goal-oriented practice appear to be often used interchangeably although their training target is different. While task-specific practice focuses on the acquisition of a specific skill for ADL, goal-oriented practice permits the use of any movement or skill that is deemed suitable to achieve the goal, fostering the exploration of alternative movement patterns. Variability appears to be included inherently in many protocols, possibly because it renders the training less repetitive and, therefore, less boring, which could counteract low adherence. However, this link has not been explicitly studied. Increasing the difficulty during practice is part of many intervention protocols as well; however, personalizing the difficulty level in order to provide training at the optimal challenge point seems to be rarely addressed. Concerning multisensory integration, it would be interesting to explore whether the presence of more than two sensory stimulations could enhance learning. Similarly, rhythmic entrainment could be extended with protocols exploring if visual or haptic entrainment might aid recovery of impaired movements. Explicit feedback and implicit feedback are often investigated together under the umbrella term of augmented feedback, as evidenced by the sourced meta-analysis and clinical studies. However, their aim and the neuronal mechanisms that they appear to stimulate are different. While explicit feedback provides terminal feedback about movement outcome, implicit feedback provides concurrent error-signals during movement execution fostering possibly different learning mechanisms. Meta-analyses also appear to interpret the sensory modality of the feedback, e.g., if it is visual, auditory or haptic as a feedback type. However, the sensory modality is a separate layer that is added to feedback. Explicit feedback, as well as implicit feedback, can be unisensory or multisensory. Action observation and mirror-therapy appear to be well studied therapeutic ingredients, whereas mental practice is only addressed in a few studies, and social interaction remains unexplored territory so far. If the principles would be better operationalized, it would not only help to identify their contribution to the recovery of motor functions, but also other learning outcomes such as cognitive or language improvements.

59The neuronal changes found within each principle allow us to draw some general conclusions for the advancement of neurorehabilitation. While some principles appear to modulate more specific brain areas (massed practice, dosage, variable practice, task-specific practice, modulate effector selection, multisensory stimulation) within the motor areas of the cortex others appear to recruit or rely more on networks of brain regions (goal-oriented practice, increasing difficulty, action observation, motor imagery, mirror therapy, rhythmic cueing, implicit feedback/knowledge of results, social interaction). An effective rehabilitation approach should thus incorporate principles of both types in order to counteract neuronal degradation and promote improvement. Firstly, a training that addresses only a limited subset of the neuronal circuitry underlying a general function might limit transfer to other behaviors that depend on the same circuitry. Secondly, not all principles are equally applicable to all patients. Some principles might be more beneficial early after stroke, whereas others benefit patients with less severe damage. Spontaneous biological recovery and activity-dependent plasticity appear to interact differently at different stages after stroke, which, aside from other factors like severity, predicts recovery. It seems that in acute patients the sensorimotor cortex activity is highly abnormal, and the normalization in activity patterns is linked to better recovery. Principles like task-oriented practice that promote localized changes, might therefore be more beneficial at the acute stage after stroke, whereas therapies like CIMT, where the forced use of the impaired limb is paired with increasing difficulty and further principles, have been shown to be more suitable at later stages after stroke and for less impaired patients. More severely impaired patients, on the other hand, might benefit from action observation, mirror therapy and motor imagery. Future studies will show the optimal combinations of principles that stimulate plasticity in a way that learning of preexisting or novel functions is enhanced.

60We are aware that the view proposed here is strongly influenced by knowledge mainly derived from clinical work with hemiparetic stroke patients. However, the literature indicates that other diseases, for instance, PD or Alzheimer’s disease, show similar cognitive, functional, and neuronal alterations even though they may have different pathologies. Therefore, these principles of neurorehabilitation could be potentially applied beyond the field of stroke. As our main goal was to provide a synthesis that is informative and practical, in-depth analysis of each principle and its neurological underpinnings lie outside of the scope of this work. In future work, we will unify the principles addressed here in a theoretical framework to show how each of them contributes to the restoration of sensorimotor contingencies.

61In summary, our review provides a synthesis of effective therapeutic ingredients that could be beneficial in aiding recovery after stroke. We hope that future work will extend the evidence presented here by implementing and investigating the principles of neurorehabilitation in novel rehabilitation protocols for stroke and other patient populations.

62MM, BB, and PV contributed to the design of the review. MM wrote the first draft of the manuscript and performed the computerized search. All authors contributed to the manuscript revision, read and approved the submitted version.

63PV is the founder and interim CEO of Eodyne Systems S.L., which aims at bringing scientifically validated neurorehabilitation technology to society. The remaining 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.