Does 70% of What Was Lost Come Back? — Stroke's "Proportional Recovery Rule" and the Illusion Statistics Built

Why This Debate

There is a person who can no longer use one arm due to a stroke. For how much this person will improve in the future, a very simple answer is available. It is about 70% of what was lost. The proportional recovery rule is a rule that a set percentage (about 70%) of lost function returns by itself. It is the answer that if you just measure how much was lost at first, the future becomes visible.

Two articles from Rehabilitation Basics have already encountered this rule. "When Does Recovery Happen — What Time Did, and What Treatment Did" used this rule as background to show how large the recovery is that happens by itself, regardless of treatment. "How Deep Is the Recovery — Between Compensation and True Recovery" went so far as to say that we cannot tell an individual patient "you will recover 70%" using this rule. Both articles put off whether the rule itself is correct until this article. Even if you have not read those two articles, you can follow this article from the beginning.

This rule is currently being refuted, and the core of the refutation lies on the subtraction side rather than the brain. The degree of improvement is the value obtained by subtracting the initial score from the later score. When this value is placed side by side with the initial score, a relationship unrelated to recovery is created purely by calculation. If the scores have a ceiling, that relationship is inflated even more. In any field that uses sentences like "the worse the beginning, the more you improve," this structure is present.

This article separates what came from the patient and what came from the calculation among the regularities shown by the rule. It also notes what the critical side does not yet know.

The Rule of Thumb: About 70% of What Was Lost Returns

The starting point for the rule is the arms of 41 ischemic stroke patients in 2008 (PMID 17687024). Motor impairment in the arms was measured using the Upper Extremity Fugl-Meyer Assessment. The Upper Extremity Fugl-Meyer Assessment (FMA-UE) is a test that scores arm movement item by item, with a maximum of 66 points and lower scores indicating greater impairment. It was measured twice: 24 to 72 hours after onset and 3 to 6 months later, and the amount of improvement in between was predicted using clinical variables such as age, sex, lesion location, infarct volume, and initial score.

The clinical variables explained only 47% of the variance in recovery. The explained variance is the portion of the differences in recovery amounts among individuals that can be accounted for by the variables used in the prediction. This means more than half was not explained by these variables.

The rule emerged from the next step. Among the patients with the most severe initial impairment, there was a separate outlier group that hardly recovered. An outlier is a value that deviates significantly from the trend drawn by the rest of the data. When they were removed, the explained variance jumped to 89%, and the recovery of the remaining patients was tightly proportional to their initial impairment. Recovery ≈ 0.70 × initial impairment (PMID 17687024).

The original scope of the rule was narrow. In mild to moderate arm paralysis, about 70% of lost upper extremity motor function returns within the first 3 to 6 months. About 30% of patients with severe initial paralysis do not follow this rule (PMID 31524062).

Why It Was Convincing — Replication, Replication, and More Replication

People who recover as much as the rule predicts are called fits, and those who recover far less than predicted are called non-fits. The two are usually separated by hierarchical cluster analysis, a statistical method that groups similar people together.

The rule emerged again in an independent group of 211 people. The prediction formula is Δpredicted = 0.7 × (66 − initial FMA-UE) + 0.4, which is 0.7 times the points missing from the maximum score, plus a small constant. Roughly 70% (146 people) showed a fixed proportional recovery of roughly 78%, and 65 people got much better less than predicted. Non-fits had more severe neurological damage within 72 hours of onset. The range where the authors noted the rule was confirmed is patients with mild or moderate initial damage (PMID 25505223).

The rule was also extended beyond the arm. In the legs, 32 people recovered 74% (95% CI 60–88%) of lower limb motor impairment between 3 days and 3 months. The 95% CI, or confidence interval, is a range of values that do not contradict these data. In this sample, a group that did not fit the rule was not seen, and the amount of treatment did not predict the recovery ratio (PMID 28341754). In 202 people, 175 people (87%) fit the leg rule, and fits got better by roughly 64% of the predicted maximum recovery. People who did not follow the rule in the legs did not follow it in the arm either (PMID 29329286).

The same shape appeared in attention, too. Hemispatial neglect is a symptom of failing to pay attention to one side of space. Eighty out of 90 people with right-hemisphere stroke showed proportional recovery in this symptom (0.97, 95% CI 0.82–1.12), and every patient who did not follow the rule in neglect also failed to follow it in the arm (PMID 27913798).

The same numbers came out in mice. Among 593 mice, roughly 30% of fits recovered 66% (95% CI 62–70%) of initial damage, and this range overlapped with the ranges of several human clinical trials (PMID 29319405).

Similar ratios came out in the arms, legs, attention, and even mice, and people who deviated from the rule in the legs or attention also deviated in the arm. Therefore, this regularity was interpreted as a common biological mechanism. Regarding the result that the amount of treatment does not affect proportional recovery, the lower limb study wrote that it is "further evidence reflecting an underlying biological process" (PMID 28341754), and the mouse study wrote that it is a "cross-species phenomenon" (PMID 29319405).

The authors of the mouse study also wrote heavier implications. It means that current rehabilitation practices may play little or no role in brain recovery, and that recovery is largely the result of natural biological recovery processes (PMID 29319405). These results and implications were covered in detail in "When Does Recovery Happen — What Time Did, and What Treatment Did". A single number even touched upon the question of what rehabilitation does. However, this implication stands on the premise that the rule is correct, and that very premise is what the rebuttal below targets.

Twist ①: The Horizontal Axis Is Hidden Inside the Vertical Axis

In 2021, a paper appeared that targeted the very foundation of the rule. Its title is its conclusion: "Inflated estimates of proportional stroke recovery: the perils of mathematical coupling and ceiling compression" (PMID 33827246).

What has been used as evidence for proportional recovery is a strong negative correlation between the initial score and the subsequent change. The change is the later score minus the initial score, and a negative correlation means that the larger one is, the smaller the other tends to be. The observation that those who lost more initially regain more is precisely this correlation. The crux of the criticism is that this correlation is already confounded in the very situation used to evaluate it (PMID 33827246). Being confounded means that causes unrelated to recovery are mixed into this correlation.

In the primer Numbers from Scratch 「If People Who Take It Live Longer, How Were They Different to Begin With? — Confounding」, confounding was when an external third factor made a relationship look skewed. Here, the factor making the relationship skewed is not an external factor, but the calculation method itself. Setting stroke aside for a moment and looking at test scores reveals this structure.

There are students who took an exam twice. Let the first-round score be x, the second-round score be y, and define the improvement as Δ = y − x. To see "whether those who did poorly at first improve more," Δ is plotted against x.

But the x on the horizontal axis is also contained inside the vertical axis. Since Δ = y − x, a large x alone pulls Δ in the direction of becoming smaller. Students who scored low on the first round are likely to show a large difference regardless of their second-round score, while students who scored high on the first round are likely to show a small difference. Even if y is determined completely unrelated to x, the slope tilts negatively and the correlation appears strong. In this graph, x is both on the horizontal axis and inside the subtraction on the vertical axis.

Mathematical coupling is a correlation that arises automatically because two axes share the same value. What this coupling inflates is the strength of the correlation rather than the value of the slope. Even if the real relationship is weak, the relationship looks stronger than it is, as if the explained variance reaches 80 to 90%.

When measurement error enters, an effect in the same direction happens once more. A person who happened to score low on the first measurement tends to go up on the next measurement as that chance occurrence disappears. Improvement created this way belongs neither to treatment nor to biology.

In stroke data, this structure appears with a more plausible face. The greater the initial impairment, the more is lost, and the more is lost, the greater the room for recovery. The observation that "those with more room to regain regain more" may not be a discovery about recovery in itself, but a statement about subtraction. The target aimed at by the criticism is not recovery itself, but the correlation used as evidence of recovery.

Twist ②: The Ceiling Makes the Rule

The second problem lies in the ruler. Ceiling compression is a phenomenon where people closer to the top of a scale have nowhere left to go, so their change is squashed. The maximum score on the FMA-UE is 66, and there is nothing above it. That is why 66 is plugged into the prediction formula earlier.

A patient whose initial score is 60 has a maximum possible recovery of 6 points. When this patient actually reaches 66, their proportion of recovery is recorded as 100%. Mild patients mostly climb near the ceiling, and because the room for recovery itself is small, the gap between the predicted and actual scores comes out small. Therefore, they are easily classified as fitting the rule, or "inliers". Severe patients cannot touch the ceiling to begin with. The pattern where the rule fits well near the ceiling and fails near the floor can arise from the scale alone. The observation that "most non-fitters are severe" may be another face of this pattern.

A cohort in Korea measured this problem directly. They screened 10,636 people and analyzed 849 patients with first-ever ischemic stroke. The ceiling effect of the FMA-UE stood out in patients with initially mild arm paralysis. When patients who differed only in their initial impairment level were paired and compared, their recovery proportions split at 0.92±0.20 versus 0.81±0.31 (P<0.001). This means the proportion of recovery was not a fixed constant, but varied depending on the initial impairment. The authors wrote that these results suggest the proportional recovery model is not valid (PMID 34134508).

An even more uncomfortable result came from a simulation. Based on FMA-UE data from 373 people, they ran virtual data engineered so that every single patient recovered completely uniformly. In this data, the very things that have been presented as evidence for proportional recovery still appeared: both the group-level regression slope (≈0.7) and the individual-level classification of fitters and non-fitters. The regression slope shows how many points larger recovery is on average when initial impairment is 1 point greater. The authors' conclusion is that commonly used methods cannot distinguish whether proportional recovery is true or not. They wrote that the boundaries of the scale can create an illusion of proportionality, and post hoc classification, such as stripping away non-fitters, can exacerbate the problem (PMID 33317423).

Post hoc classification connects back to the starting point of the rule. The initial observation in 2008 transparently reported that explanatory power went up after removing outliers (PMID 17687024). What the simulation showed are two things: that even in data engineered so everyone recovers by the same amount—that is, non-proportional data—a group of non-fitters and a slope near 0.7 emerge, and that post hoc classification can inflate this problem (PMID 33317423). If you remove people who do not fit first and then re-measure the goodness of fit on the remainder, the fit is bound to go up.

The parts of the illusion are thus three: the initial score shared by both axes, the ceiling of the scale, and post hoc classification. None of the three says there is no recovery. Yet when the three overlap, it becomes impossible to separate how much of the observed regularity came from the patient and how much came from the ruler and subtraction.

Twist ③: When You Measure Again Avoiding Confounding, Explanatory Power Shrinks

Then what remains when you measure again while avoiding confounding? Recovery trajectories after stroke for 385 people gathered from six studies attempted to answer this question. It belongs to the largest scale for upper limb motor recovery data. The research team used a hierarchical Bayesian model that puts multiple datasets together into one model for estimation, and handled the confounding of the ceiling effect with a subset approach. After checking if the estimation was correct through synthetic data simulations, they compared the models (PMID 32601678).

The results came out in two stages. When looking only at fitters as customary, recovery was best explained by "a component proportional to what was lost + a constant component." The direction is the same as the rule. However, the recovery variance explained by this model was 32%. This is very different from the 80% or more that has been reported previously.

When looking at the whole by including both fitters and non-fitters, the winning model changed. "A component proportional to what remains + a constant" was preferred. Rather than the side that recovers more the more it lost at first, it is the side that recovers more the larger the remaining function is. The explained variance was 53% (PMID 32601678).

This result must be carried over with subtlety. The authors wrote that motor recovery after stroke may show some characteristics of proportionality. However, what it is proportional to was flipped, and the explanatory power decreased significantly compared to previous reports. In the analysis looking only at fitters, it fell short of even half of previous reports. The authors concluded that behavioral scores alone are not enough to explain stroke recovery, and that research beyond that is necessary to establish robust and discriminative single-patient predictions (PMID 32601678).

Then What Was Real — Let's Set Up the Other Side Too

There is evidence and logic on the rule side, and the critics also have points that are not yet settled.

The most organized counterargument from the rule side is a 2019 methodological paper. Group-level models are models that describe the average recovery of a group of patients instead of the recovery of a single patient. After closely comparing studies that reported proportional recovery and examining statistical techniques, this paper concluded that existing data largely match the rule as a group-level model of upper-limb motor recovery. The recent reports announcing the end of the rule were exaggerated, and those reports focus too much on individual patient-level prediction, which is a less conclusive problem (PMID 31524062).

The same paper acknowledges its weaknesses. As the rule expanded to other impairment domains, formulation and application became inconsistent across studies, making it difficult to draw conclusions across studies (PMID 31524062).

The critics replied back to this position. The earlier 2019 paper (PMID 31524062) reevaluated the classical rule and viewed that "effectively predicting an individual patient's recovery trajectory is not supported by current evidence, but stating the existence of proportional recovery at the group level is reliable," and a 2021 critical paper countered this reevaluation by stating that the confounding is identically present in both individual-level and group-level inferences (PMID 33827246). The correlation between the initial score and the amount of change is subject to the same subtraction and the same ceiling effects whether you look at patients one by one or as a group.

Therefore, the dispute lies in group-level interpretation, and neither side supports individual prediction with this rule. This is the retort briefly delivered by "How Deep Is the Recovery — Between Compensation and True Recovery".

Even After Handling Confounding, the Subgroups Remained

Data from 489 people gathered across eight cohorts handled severe and non-severe cases simultaneously within a single model. To avoid the confounding of previous recovery studies, an explanatory model was built using only patients with an initial FMA of less than 45 points, and the boundary separating severe and non-severe cases was explored sequentially within the initial FMA range of 5 to 30 points.

The data fit best to a model assuming two subgroups, and the boundary was an initial FMA of 10 points. The two subgroups recovered similar sizes of the constant component, but the directions of the proportional component differed from each other. In the severe group, those with smaller impairment recovered more, whereas in the non-severe group, those with larger initial impairment recovered more. The predictive performance for the 3- to 6-month outcomes was R²=63.5% (95% CI 51.4–75.5%).

The authors concluded that dividing severe and non-severe cases in the recovery model is not an artifact of previous confounding (PMID 34937750). Even if the label of non-responder may have been a product of post hoc classification, the very fact that severe patients trace a different trajectory survived. There are also findings favorable to the rule. Even after handling confounding, the non-severe group retained the same direction as the rule, where larger initial impairment led to more recovery.

Instead of a Single Straight Line, Multiple Trajectories

When the recovery of 412 first-time ischemic stroke patients is plotted as a curve that flattens over time, five distinct groups emerge. Each group has its own proportionality constant and time constant, which indicates how quickly recovery tapers off. A model that assumes multiple groups are mixed within the data and sets out to find them is called a mixture model. This model started from the criticism that the 70% rule overestimates the predictability of FMA-UE recovery (PMID 31925838).

The median absolute error is the middle value when the differences between predictions and actual outcomes are lined up by magnitude alone. The median absolute error of the final FMA-UE was 4.8 points (IQR 1.3–12.8) at 1 week post-stroke and 4.2 points (IQR 1.3–9.8) at 2 weeks. The overall accuracy of assigning recovery groups (poor, intermediate, good) was 0.79 at 1 week and 0.81 at 2 weeks (PMID 31925838).

The IQR, or interquartile range, is the span where the middle half of the people lie. The median error is within 5 points, but the upper end of this span is 12.8 points at the 1-week mark and 9.8 points at 2 weeks. This means most predictions are close while some are far off. The authors concluded that final scores and recovery groups can be predicted well (PMID 31925838). Changing the rule from a single-line straight line to a bundle of multiple trajectories improves the explanation, but the uncertainty in front of any individual patient does not disappear.

The same data can become two stories depending on how you look at it. 201 WAKE-UP trial patients who had moderate impairment at onset and residual impairment at 22 to 36 hours were repeatedly grouped based on how similar their NIHSS recovery ratios were to one another. The NIHSS is a score that rates the severity of stroke symptoms, where higher means more severe. There were six groups. Complete recovery, above average, two average groups that fit the rule, below average, and worsening (PMID 41612435).

Fitting a straight line to these same patients yields the same result as the literature, namely the adapters' average recovery ratio of 70%. Yet for most patients, the NIHSS did not decrease proportionally. Lesion volume did not significantly differ between the groups, and left-hemisphere stroke was more common in the group with lower recovery (PMID 41612435). The same data becomes the "70% rule" when viewed as a straight line, and six different courses when viewed as groups.

The Rule Is Still Used as a Comparison Standard

In 2026, electroencephalogram (EEG)-based machine learning models report their performance side by side with the proportional recovery rule. In subacute FMA-UE prediction, the median absolute error was 5.85 for the model and 19.00 for the rule. In the acute sample, the error was 5.87 for the model trained with subacute data combined and 8.80 for the rule. The model trained only on acute data had an error of 13.74, which was larger than the rule (PMID 41641382). This means that in the acute stage, subacute data had to be added for the model to outperform the rule, and that this rule remains the benchmark to beat today.

What Both Sides Agree on Together

What the critics established is "it cannot be judged with existing evidence," and they did not go as far as to say "proportional recovery does not exist." Analyses that re-measured while avoiding confounding have also produced results showing that recovery can exhibit some proportional characteristics (PMID 32601678). Even the conclusion of the paper at the peak of the criticism is that a new analytical technique is needed, and it does not put forward an alternative model (PMID 33827246). The rule side also wrote that methodological caution and new analytical approaches are needed to confirm or refute the systematic nature of spontaneous recovery, whether at the group level or the individual level (PMID 31524062).

What both sides agree on together goes only as far as saying that a new analysis method is needed. At the same time, the rule side maintains the judgment that existing data generally match group-level models, while the critic side views that judgment as also standing on the same confounding. The current state is closer to "it cannot be judged with current tools" than to "the rule is dead."

Where Did Clinical Prediction Go — Algorithms and Biomarkers

Apart from the rules debate, the clinic needed answers. The answer moved away from a single behavioral score toward algorithms and biomarkers. Biomarkers are indicators like imaging or nerve tests that measure the state inside the body to help gauge prognosis.

PREP2 is a prediction algorithm built from data on 207 patients within 3 days of onset. It sequentially combines an arm impairment score (SAFE, shoulder abduction and finger extension muscle strength), age, the presence of motor evoked potentials (MEPs), and MRI lesion load or the NIHSS to predict 3-month upper limb function into four categories. MEPs are the responses captured in arm muscles when magnetic stimulation from outside the head touches the motor area of the brain. In the development cohort, predictions were correct for 75% of patients (PMID 29159193).

In external validation, the authors reported very good predictive power. In a prospective cohort of 143 patients, the overall accuracy was 78% (95% CI 71.2–86.1%). However, the variance across categories was large. Accurate predictions were 87.5% for EXCELLENT and 94.9% for POOR, while GOOD was 46.5% and LIMITED was 0%. In patients with cognitive impairment showing a score below 24 on the cognitive screening test MoCA, the accuracy was 69.4%, which was lower than the 93.1% in patients without cognitive impairment (P=.01) (PMID 39162251).

The two poles fit well, but the middle does not fit well. And in the clinic, the patients whose prognoses are most curious are usually in the middle.

In severely affected patients where the rules deviated the most, information on the corticospinal tract shows its usefulness. The corticospinal tract is a nerve bundle that sends movement instructions straight from the motor area of the brain down to the spinal cord.

• Weighted corticospinal tract lesion load (wCST-LL) calculated from acute-phase imaging predicted 3-month impairment as well as the initial clinical assessment. In the severe subgroup with an initial UE-FM of 10 points or less, it significantly outperformed clinical assessment (R²=0.47 vs. 0.11, p=0.03). The non-severe group recovered about 70% of their maximum recovery potential (PMID 26289123).
• The presence or absence of MEPs points in the same direction. Among 61 patients with moderate-to-severe acute ischemic stroke, patients with preserved MEPs showed better recovery (about 70%) at 90 days. For patients with an initial FMA above 38 points, proportional recovery was 65.2±27.0% at 30 days and 83.7±24.6% at 90 days (PMID 37345693).

The standard deviation following the ± in the last two values shows how scattered individual recovery rates are from the average, and that spread is wide. That the group average is near 70% and that an individual recovers 70% are different statements.

Significance: This Story Is Bigger Than Stroke

What this debate leaves behind is how to handle the amount of change, rather than the fate of the number 70. The three parts we saw earlier—the initial value and the ceiling of the scale split between two axes, and post-hoc classification—do not belong to rehabilitation medicine alone. When diet research says, "the higher the initial weight, the more you lose," when education research says, "the lower the initial grades, the more they rise," and when corporate research says, "teams that start smaller grow larger," the same parts are at work. Before a sentence that says, "the worse the start, the more it improves," we must first ask whether it is a discovery or subtraction.

There is also credit due on the side of the rules. The proportional recovery rule placed the question "is there a structure to recovery?" right in the middle of this field, and the mixed models and new prediction models we saw earlier all emerged by being compared against this rule. Even a flawed map is sometimes better than having no map at all.

The sentence to use in the clinic can be summed up in one. As "How Deep Is the Recovery — Between Compensation and True Recovery" noted, you cannot use this rule to say, "this patient will recover 70%." A sentence that may be true for a group is not necessarily true for an individual. That gap is why this debate is not yet over.

Source Papers

• Inflated Estimates of Proportional Recovery From Stroke: The Dangers of Mathematical Coupling and Compression to Ceiling. PMID 33827246 (2021) — The correlation between initial scores and change is confounded, and that confounding exists at both the individual and group levels. Concludes that new analysis techniques are needed.
• Statistical Considerations for Drawing Conclusions About Recovery. PMID 33317423 (2021) — Simulation based on n=373 data. Even in uniformly recovering data, a slope ≈0.7 and the classification of fitters/non-fitters are reproduced.
• Revisiting the Proportional Recovery Model in View of the Ceiling Effect of Fugl-Meyer Assessment. PMID 34134508 (2021) — 849 people. After propensity score matching, proportional recovery was 0.92±0.20 versus 0.81±0.31 (P<0.001). "Suggests the model is not valid".
• Bringing proportional recovery into proportion: Bayesian modelling of post-stroke motor impairment. PMID 32601678 (2020) — 385 trajectories. Looking only at fitters, explained variance is 32%; across the full spectrum, 'proportional to what remains + constant' is preferred at 53%.
• What the Proportional Recovery Rule Is (and Is Not): Methodological and Statistical Considerations. PMID 31524062 (2019) — As a group-level model, it largely matches the data, but "endpoint reporting is exaggerated." Individual-level prediction is the weak link in the debate.
• Inter-individual variability in the capacity for motor recovery after ischemic stroke. PMID 17687024 (2008) — The starting point of the rule. 41 people, clinical variables explain 47% of variance → 89% after removing outliers, recovery ≈ 0.70 × initial impairment.
• Generalizability of the Proportional Recovery Model for the Upper Extremity After an Ischemic Stroke. PMID 25505223 (2015) — 211 people. About 70% (146 people) show fixed proportional recovery of about 78%, and 65 people are non-fitters. Prediction formula 0.7×(66−initial)+0.4.
• Recovery after stroke: the severely impaired are a distinct group. PMID 34937750 (2022) — 489 people, 8 cohorts. Two subgroups at the boundary initial FMA score of 10, prediction R²=63.5%. The distinction between severe/non-severe is not an artifact of confounding.
• Predicting Upper Limb Motor Impairment Recovery after Stroke: A Mixture Model. PMID 31925838 (2020) — 412 people, 5 subgroups. Median absolute error of final FMA-UE was 4.8 at 1 week and 4.2 at 2 weeks, with a cluster assignment accuracy of 0.79 and 0.81.
• Beyond proportional recovery in wake-up stroke: unsupervised recovery clusters based on the NIHSS. PMID 41612435 (2026) — 201 WAKE-UP patients. Linear fit reproduces 70%, but NIHSS does not decrease proportionally in most patients. 6 clusters.
• Proportional Recovery From Lower Limb Motor Impairment After Stroke. PMID 28341754 (2017) — 32 people, lower limb 74% (95% CI 60–88%). No non-fitter cluster, and treatment dose is not a predictor.
• Is the proportional recovery rule applicable to the lower limb after a first-ever ischemic stroke? PMID 29329286 (2018) — Out of 202 people, 175 (87%) fit, and the average improvement of fitters is about 64% of the predicted maximum. Lower limb non-fitters are also non-fitters in the upper limb.
• Generalizability of the Maximum Proportional Recovery Rule to Visuospatial Neglect Early Poststroke. PMID 27913798 (2017) — Out of 90 people, 80 fit (0.97, 95% CI 0.82–1.12). All neglect non-fitters are also non-fitters in the upper limb.
• Does Stroke Rehabilitation Really Matter? Part A: Proportional Stroke Recovery in the Rat. PMID 29319405 (2018) — 593 rats, about 30% are fitters, recovery 66% (95% CI 62–70%). Clarifies the implications the rule has for the role of rehabilitation.
• PREP2: A biomarker-based algorithm for predicting upper limb function after stroke. PMID 29159193 (2017) — 207 people. Sequentially combines SAFE, age, MEP, and lesion load/NIHSS, with 75% accuracy.
• Prediction of Upper Limb Motor Recovery by the PREP2 Algorithm in a Nonselected Population: External Validation and Influence of Cognitive Syndromes. PMID 39162251 (2024) — 143 people. Overall 78%, but GOOD is 46.5% and LIMITED is 0%, with 69.4% versus 93.1% in MoCA<24 (P=.01).
• Corticospinal tract lesion load: An imaging biomarker for stroke motor outcomes. PMID 26289123 (2015) — 76 people. In the severely impaired group, wCST-LL outperforms clinical assessment (R²=0.47 vs. 0.11, p=0.03). Non-severely impaired groups recover about 70% of maximum potential.
• Predictive value of motor-evoked potentials for motor recovery in patients with hemiparesis secondary to acute ischemic stroke. PMID 37345693 (2023) — 61 people. MEP-positive group recovers about 70% at 90 days. Proportional recovery in the initial FMA>38 group is 65.2±27.0% at 30 days and 83.7±24.6% at 90 days.
• Enhancing upper limb motor recovery prediction after acute stroke using EEG and subacute data. PMID 41641382 (2026) — Uses the proportional recovery rule as a benchmark. Subacute median absolute error 5.85 vs. 19.00, acute 5.87 vs. 8.80.