How Deep Is the Recovery — Between Compensation and True Recovery

When you look side by side at two patients who have picked up a cup again, a difference shows. One reaches the hand out by straightening the elbow. The other brings the hand to the cup by leaning the torso forward with the elbow left almost as it is. Both succeeded at the task, and their scores can go up together. Yet what happened inside the body is different.

The previous article, "How Do We Measure Recovery — An Introduction to Rehabilitation Outcome Measures", covered how to read outcome measures. This article deals with the question left behind after the scores go up. It asks: when the score has risen, how do we tell apart whether the arm has returned or whether one has learned a way to do it differently?

The meanings of the two terms were established in the primer piece, "Neuroplasticity — A Brain That Builds New Roads Beside the Broken Ones". Recovery is when movement or function returns in the same way as before the injury, and compensation is accomplishing the same thing by using another part of the body or a different method instead of the lost way. Reaching the cup by tilting the torso is compensation seen through the shape of movement. That piece briefly introduced up to the point that the two are hard to separate, and that a proposal recently emerged to ask again with "how deep is it". This article is the step after that. It follows how the two are actually measured, why compensation is not only bad, and how this field is rebuilding that distinction itself. Improvement in the broad sense is written as "amelioration" in this article.

Why We Need to Know This — The Same Score, Different Events

This distinction matters in rehabilitation practice for three reasons:

• We can misread treatment effects. We might increase compensation alone and conclude that the patient has "recovered."
• Long-term outcomes can differ. There has long been concern that relying on trunk compensation builds up strain on the shoulder and core.
• What we train changes. Whether to suppress or permit compensation branches from here.

The Trunk Reaches Instead, Even in Everyday Tasks

Compensation is not a phenomenon seen only in controlled tasks in the laboratory. It also appears in tasks close to everyday life.

When using the paralyzed arm, the trunk moved more and the elbow moved less. When nine chronic stroke survivors performed a sequence of reaching and manipulation tasks, the movement and rotation of the trunk were significantly greater when they used the paralyzed arm (paretic side) compared to their own healthy arm (non-paretic side) ($P = .03$). Conversely, the change in elbow angle was significantly smaller ($P = .01$). The task of reaching for the hand required more trunk compensation and elbow movement than the task of handling objects. For comparison, 20 control groups without neurological diseases also performed the same task (PMID 33631718).

The hand reaches the target, but the trunk fills that distance instead of the elbow. The authors saw this reliance on trunk compensation even in everyday tasks as a target to be aimed at by home-based rehabilitation (PMID 33631718). However, this is an observation from a sample with a small number of patients.

Scores Look at the Outcome — To See the Process, You Need Kinematics

Clinical metrics usually check whether a task was accomplished, meaning they look at the outcome. Therefore, they cannot distinguish whether something was done with the trunk or with the arm. To tell the two apart, you have to look at the movement itself. Kinematics is the measurement of when, in what direction, and at what speed joints moved. It is obtained by attaching markers to the body and recording the movement with cameras or sensors.

Translated into the words of "Rehabilitation — Starting to Plan a Livable Home as the Fire Is Brought Under Control" from the primer, whether a task was accomplished is a question about the level of activity, and how the joints moved is a question about the level of body function. Different body movements can hide beneath the same activity score.

Two Groups Under the Same Improvement — Exoskeleton Training Research

Kinesiology's clearest example of actually separating recovery from compensation is exoskeleton training research. An exoskeleton is a mechanical frame strapped to the arm that supports the arm's weight and records movement. 53 patients in the subacute stage—the weeks to months after passing the initial crisis of a stroke—performed about 80 reaching movement tests each week for 4 weeks while wearing an exoskeleton called ArmeoSpring. All participants showed improvement in distal performance. Distal performance looked at how smooth the movement of the end of the arm, namely the hand, was. Up to this point, it looks as though everyone recovered (PMID 33994981).

The researchers looked at joints instead of the hand. They analyzed four angles together: shoulder horizontal rotation, shoulder elevation, elbow rotation, and forearm rotation. The benchmark was a healthy young control group. In their reaches, two patterns of joints moving in pairs accounted for more than 95% of the changes in joint velocity. One is the pairing of shoulder horizontal rotation and the elbow, and the other is the pairing of shoulder elevation and forearm rotation. This means that reaching with a healthy arm consists mostly of these two pairs (PMID 33994981).

Following how these two pairs changed during training divided the participants into two groups. There were 19 people in the "recovery group," where both pairs moved closer to healthy values. In the remaining group, at least one pair moved further away from the healthy values. They figured out new compensation patterns different from healthy people and recreated smooth hand movements. The authors divided the two groups into "recovery" and "compensation," and wrote that this analysis could be used to refine exoskeleton therapy to provide movement feedback during training and reduce compensation patterns (PMID 33994981).

It looked like the same improvement, but joint coordination returned toward normal in only 19 out of 53 people. If we had only looked at the outcome of hand smoothness, the two groups would have been indistinguishable.

Compensation Is Not Always Bad

If we read compensation only as a sign of failure, we miss something. People who have recovered well also use compensation.

Thirteen stroke patients who had recovered well and 13 control participants matched for age, sex, and dominant hand each reached out and grasped objects freely 80 times. Movements were recorded at 200 Hz using optical motion capture. The researchers divided the movements into hand movement (distal kinematics) and the movement of joints closer to the torso like the shoulder and elbow (proximal kinematics). Hand movement had returned to the point of being nearly indistinguishable from the control group. The difference lay in the timing of when the shoulders and elbows moved. The authors concluded that these results point to compensatory coordination of the shoulder and elbow to achieve functional goals like reaching (PMID 40525283). This is a cross-sectional study comparing two groups at a single point in time.

This result makes us read compensation differently. The arm has redundant degrees of freedom. This means there are multiple combinations of joint angles that bring the hand to the same spot. The proximal compensation of well-recovered individuals may be a way of choosing a pathway they can use stably among those multiple options.

There are also actual measurements showing why compensation arises. Force reserve is the share of force the shoulder can produce that is left over rather than fully used. Twenty healthy adults were made to lift increasingly heavy dumbbells while reaching to simulate a situation where shoulder strength falls short. As the dumbbells got heavier, they moved their torso further forward, and the force reserve predicted the occurrence of this compensation. The authors interpreted compensatory movement as a strategic use to conserve weakened shoulder strength and distribute the load across multiple joints. Translating this to compensation in stroke patients is an implication the authors raised, and this experiment did not measure it directly from patients (PMID 38985941). The compensation shown by this experiment is a solution the body finds when reserves are lacking.

There is a reverse concern as well. It is that trunk compensation can block functional recovery. The devices to suppress compensation that we will see later started from this concern (PMID 38894119, PMID 42464320). When compensation helps and when it gets in the way have not been settled into a single conclusion.

Instead of "Is It Real?", "How Deep Is It?" — The Stratified Model

A reorganization proposed in 2026 comes closest to the conclusion of this article. The starting point of this paper is this: The very attempt to sharply separate "true" recovery and compensation actually obscures the nature of motor improvement in neurorehabilitation (PMID 41877523).

What has been proposed as an alternative is the stratified model. The stratified model does not divide motor improvement into two categories of recovery or compensation, but is a framework that defines it by depth—specifically, how far down it has reached among the multiple layers of the motor system. This framework views the motor system as a hierarchy of causally linked layers. On the inside is a physiological layer such as synaptic plasticity, and on the outside is a behavioral layer such as task performance. Improvement can happen at any layer, and appears as a compensation-recovery gradient propagating from the inside outward (PMID 41877523). This means that a state where only the outer behavioral layer has improved is close to the compensation end of the gradient, while a state where even the inner layers have improved together is close to the recovery end.

The core proposition of this model is this: The question is not whether recovery is "real," but how deep it is. When behavioral-level improvements are supported by positive changes in the inner physiological layer, the authors view that as a more complete and desirable clinical outcome. When comparing two treatments with the same behavioral outcomes, the one that produces recovery at a deeper layer as well is given priority. They wrote that to verify depth, clinical scales alone are not enough, and movement analysis measured with instrumentation and neurophysiological measurements must be used together (PMID 41877523).

The authors of the stratified model did not deal with the exoskeleton study discussed earlier. However, we can apply this framework to read it in our own way. If we place the smoothness of the hand as the result of the outermost layer and joint coordination one layer inward, it reads that the two groups improved equally in the outer layer, and the improvement of the recovery group, whose joint coordination even returned to the healthy side, was one layer deeper. Since the inner layer that this model speaks of is a physiological layer like synaptic plasticity and joint coordination is still a layer of movement, this is not a result measured for depth, but an interpretation borrowing the framework.

A Rule for Predicting the Magnitude of Recovery, and Its Rebuttal

A frequent question alongside depth is the magnitude of recovery. The proportional recovery rule is a rule stating that most stroke survivors recover a fixed proportion (roughly 70%) of the function they lost. A strong negative correlation between the initial score and the subsequent change—meaning that those who lost more at the beginning tend to regain more—was used as its basis.

This basis was attacked methodologically. The core critique is that the correlation between the initial score and the change is already confounded under the common circumstances in which it is evaluated. There are two causes. One is mathematical coupling. Because the change is the later score minus the initial score, the initial score enters into both axes, automatically producing a correlation unrelated to recovery. The other is compression toward the ceiling. Scales have a maximum score, and the scores of mildly injured people cluster near that ceiling, making the proportion look consistent.

Following this criticism, two re-evaluations appeared. One modeled the data while avoiding the confounded correlation, while the other re-evaluated the classic form of the rule. The latter claimed that while effectively predicting an individual patient's recovery trajectory is not supported by current evidence, saying that proportional recovery exists at the group level is reliable. The side that raised the critique rebutted this in turn. That confounding exists equally at both the individual level and the group level (PMID 33827246).

What both sides are arguing over is group-level interpretation, and neither side supports this rule for individual prediction. Therefore, you cannot tell an individual patient "you will recover 70%" using this rule. The detailed course of the debate is covered separately in Does 70% of What Was Lost Come Back? — Stroke's "Proportional Recovery Rule" and the Illusion Statistics Built.

How Different Are Clinical Metrics and Kinematic Metrics?

Then, in actual research, which metric captures changes better? Responsiveness is how well a metric captures actual changes before and after treatment. There is a systematic review and meta-analysis comparing this. The subjects were 12 randomized controlled trials (191 participants). The treatments evaluated by the trials were three: constraint-induced movement therapy (CIMT), which ties down the sound arm and makes the paralyzed arm used intensively; trunk restraint, which fixes the torso to a chair to reach with the arm; and bilateral upper extremity therapy using both arms. As kinematic metrics, smoothness, movement time and efficiency, and ranges of motion of the trunk and shoulder were reported (PMID 32156646).

It is easy to expect that kinematic metrics, which directly measure movement, would be more sensitive to changes. The result was the opposite. The one with the highest responsiveness was the clinical metric Motor Activity Log. This questionnaire asks how much and how well patients used their paralyzed arm in daily life. The standardized mean difference (SMD), a value that matches effect sizes to a unit of dispersion, was 1.0 for amount of use and 0.96 for quality of movement. Kinematic metrics such as movement efficiency, ranges of motion of the trunk and shoulder, control strategy, and maximum velocity were lower than that. The authors concluded that this result aligns with existing literature stating that using kinematic and clinical metrics together makes for an accurate evaluation (PMID 32156646).

The two metrics answer different questions. Clinical metrics capture well whether a change occurred, and kinematic metrics show what kind of movement that change was achieved with. The stratification model earlier suggesting to add motion analysis to clinical scales is also in the same direction.

Attempts to Reduce Compensation

Attempts to actually reduce or catch compensation, especially trunk compensation, look like this:

• An approach using visual feedback to reduce trunk compensation during robot-assisted reaching exercises (PMID 38894119).
• An approach suppressing trunk compensation during unsupervised, solo upper limb rehabilitation using smartphone-based real-time feedback (PMID 42464320).
• An approach identifying compensation strategies by analyzing upper body kinematics with inertial sensors (PMID 41471605), alongside attempts to detect compensation using a standard webcam and deep learning (PMID 41322226).

They share a single commonality. Compensation is only caught by directly looking at movement, which is why measurement tools are moving beyond clinical indicators.

Common Misconceptions

The idea that "compensation is bad" cannot be taken as absolute. Well-recovered patients also appeared to use proximal compensatory adjustments to achieve reaching (PMID 40525283). It is a solution that emerges when there is no strength to spare. However, these measurements are from a simulation that applied loads to healthy individuals (PMID 38985941).

The idea that "the score went up, so the patient has recovered" is refuted by exoskeleton research. All participants showed improvement in distal performance, but joint coordination converged to normal in 19 out of 53 people. Furthermore, the two groups showed no difference at either the start or the end of training in Fugl-Meyer upper extremity scores, which measure the degree of impairment. This means that different things were happening beneath the same score (PMID 33994981).

Regarding the idea that "it is either true recovery or compensation, one of the two," recent proposals suggest that this dichotomy itself creates ambiguity. Therefore, the question is re-asked as a matter of depth (PMID 41877523).

The idea that "we can state this patient's prognosis with the proportional recovery rule" is not supported by evidence. Predicting the recovery trajectory of individual patients is not supported by current evidence (PMID 33827246).

To Go Deeper

• When reading papers, check which metrics report outcome (task success) and which report process (joint coordination). Without both, compensation and recovery cannot be separated.
• When comparing treatments, the stratification model proposes that if behavioral outcomes are equal, the one that makes changes at a deeper layer should be seen as the better result (PMID 41877523).
• The balancing act between compensation achieved using the unaffected side and practice that forces the injured arm to be used is covered in "Able But Unused — An Introduction to Learned Non-Use and CIMT".
• The next piece, "One Word, "Spasticity" — Reading the Upper Motor Neuron Syndrome Apart", moves on to the symptom axis. Upper motor neurons are the nerve cells that send movement instructions down from the brain to the spinal cord, and we distinguish how spasticity, muscle weakness, and contracture differ when that pathway is injured.

Source Papers

• How deep is your recovery? Rethinking true motor recovery versus compensation in neurorehabilitation. PMID 41877523 (2026)
• Compensatory Proximal Adjustments Characterize Effective Reaching Movements After Stroke. PMID 40525283 (2025)
• Dissociating Sensorimotor Recovery and Compensation During Exoskeleton Training Following Stroke. PMID 33994981 (2021)
• Compensatory Trunk Movements in Naturalistic Reaching and Manipulation Tasks in Chronic Stroke Survivors. PMID 33631718 (2021)
• Force reserve predicts compensation in reaching movement with induced shoulder strength deficit. PMID 38985941 (2024)
• Responsiveness of kinematic and clinical measures of upper-limb motor function after stroke: A systematic review and meta-analysis. PMID 32156646 (2021)
• Inflated Estimates of Proportional Recovery From Stroke: The Dangers of Mathematical Coupling and Compression to Ceiling. PMID 33827246 (2021)
• Mitigating Trunk Compensatory Movements in Post-Stroke Survivors through Visual Feedback during Robotic-Assisted Arm Reaching Exercises. PMID 38894119 (2024)
• Smartphone-based real-time feedback to suppress trunk compensation for unsupervised upper limb rehabilitation in patients with brain disorders. PMID 42464320 (2026)
• Compensation Strategies in Post-Stroke Individuals: Insights from Upper Body Kinematics Analysis Based on Inertial Sensors. PMID 41471605 (2025)
• Using deep learning to detect upper limb compensation in individuals post-stroke using consumer-grade webcams-A feasibility study. PMID 41322226 (2025)
• Post-stroke upper limb rehabilitation: clinical practices, compensatory movements, assessment, and trends. PMID 40698522 (2025)