Rehabilitation after stroke works partly by encouraging the brain to rebuild connections between neurons, a process called synaptic plasticity. Animal work suggests the dose matters: rats whose treadmill intensity was raised gradually recovered better and had more of the growth factor BDNF than rats trained at a fixed high intensity, which raised stress hormones. Testing whether this holds in people increasingly happens at home, with wearables and apps. That shift brings three distortions at once. People without devices or digital skills enroll less and drop out more. Sensor-based scoring models are least accurate for the most impaired patients. Patients and caregivers tend to report gains that validate their effort. All three distortions track how impaired the patient is, so they compound rather than cancel. A remote trial can therefore report a benefit that is mostly an artifact of which patients were measured and how.
We propose a two-phase "anchored" trial design and an analysis that removes this compound bias by design rather than by assumption. Every participant receives cheap remote measures. A random subsample, with known sampling probabilities and deliberate oversampling of dropouts, people who never used the device, and severely impaired patients, receives a blinded in-person motor assessment. Because the sampling probabilities are set by the investigators, the gap between the remote proxy and the anchor can be estimated without any untestable assumption about why people dropped out. The aim is a cheap, unbiased estimate of whether graded-intensity home rehabilitation beats fixed-intensity rehabilitation, together with a measured breakdown of how much each bias contributes.
Core Claim
The idea of combining wearables with neuroplasticity rehabilitation is not new. Our claim is narrower: in remote rehabilitation trials, selection bias acts at two levels. The first is enrollment and retention. The second is measurement, because the scoring instrument itself drops or misreads the patients it handles worst. A small, designed in-person subsample that deliberately reaches the people the remote system lost can correct both levels, and it can also separate arm-differential self-report inflation from real change. Graded versus fixed training intensity is the test case because the biology predicts a real but modest effect, and modest effects are the ones these biases can manufacture or erase.
Why the Pieces Have Not Been Joined
Three groups of prior work motivate the design.
The plasticity evidence comes from supervised settings. Gradually Increased Training Intensity Benefits Rehabilitation Outcome after Stroke by BDNF Upregulation and Stress Suppression compared fixed and graded treadmill intensity in 60 rats over 7 days. Graded training raised hippocampal and striatal BDNF and avoided the corticosterone rise seen with fixed high intensity. Cognitive intervention combined with aerobic limb rehabilitation exercise promotes nerve and limb function recovery in hemiplegic patients with hypertensive intracerebral hemorrhage ran a 6-month in-clinic program in 86 patients and reported higher BDNF, lower S100B and myelin basic protein, and better scores on the MMSE (a standard cognitive screening test), the Fugl-Meyer Assessment (FMA, a therapist-scored motor scale) and the Barthel Index (a scale of independence in daily activities). That study used a caregiver-centered care model and unblinded scales. It supplies our intervention regime but not a bias-resistant outcome.
The selection evidence comes from mobile health. A remote measurement study of PTSD and cannabis use among veterans found that requiring devices selected for tech-proficient veterans. In that study, retention fell to 72% by month three and daily survey completion averaged 68%. Device requirements also skewed the sample in Impact of Push Notifications on Physical Activity and Sodium Intake Among Patients with Hypertension, which required smartphone ownership, and in Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns, whose sample was entirely medical interns. Understanding How Older Adults Comprehend COVID-19 Interactive Visualizations via Think-Aloud Protocol shows older users abandoning interfaces entirely. These papers motivate the claim that dropout is likely missing-not-at-random, meaning it depends on things the remote data never record. They also constrain the design: the remote arm must require almost no interaction with an interface.
The measurement evidence supplies our remote proxies and exposes their weak spot. StrokeRehab: A Benchmark Dataset for Sub-second Action Identification segments wearable motion data into sub-second movement primitives such as reach and transport. Its authors report that the model degrades for severely impaired patients. Estimation of Fugl–Meyer Assessment Upper-Extremity Sub-Scores Using a Mixup-Augmented LSTM Autoencoder and Wearable Sensor Data predicts therapist FMA upper-extremity scores from seven motions completed in under 10 minutes. It was tested on only 11 patients in a supervised setting.
Neither paper asks how the error in the score is distributed across impairment levels, or what that distribution does to an estimated treatment effect. Exploratory Hidden Markov Factor Models for Longitudinal Mobile Health Data recovers latent recovery states from wearable heart-rate-variability features. Its states are interpreted against self-report surveys, which illustrates the anchoring problem we want to fix.
Research Program
The program has three stages.
The first stage is a measurement audit on existing benchmarks. We retrain the sequence-to-sequence primitive model on StrokeRehab and the FMA-UE estimator using the published protocols. We then report error as a function of baseline impairment instead of a single pooled number. The output is an error profile: bias and variance of the remote score at each severity level.
The second stage is a semi-synthetic simulation. We generate trial populations with a true effect of graded versus fixed intensity and plug in three sources of distortion. Attrition curves come from the veterans study, at 92%, 76% and 72% retention across three months. Severity-dependent scoring error comes from stage one. Arm-differential reporting inflation is varied over a range, since the corpus offers no estimate of its size.
We then compare analysis strategies on bias, coverage and cost. For each arm $a$, the anchored estimator of mean recovery is
$$\hat\mu_a = \frac{1}{N_a}\sum_{i \in a} \tilde Y_i \;+\; \frac{1}{N_a}\sum_{i \in S_a} \frac{Y_i - \tilde Y_i}{\pi_i},$$
where $\tilde Y_i$ is participant $i$'s remote proxy score (model-imputed when the device data are missing), $Y_i$ is the blinded in-person FMA, $S_a$ is the anchored subsample in arm $a$, and $\pi_i$ is the probability, set in advance, that participant $i$ was drawn into it. The treatment effect is $\hat\mu_{\text{graded}} - \hat\mu_{\text{fixed}}$. The estimator is unbiased whenever $\pi_i > 0$ for everyone, dropouts included, and anchors are actually obtained. It follows from standard two-phase sampling, so no model of the dropout mechanism is needed. The cost is variance, which shrinks as the proxy tracks the anchor more closely. The simulation also chooses the oversampling scheme that minimizes variance per in-person visit.
The third stage is a pilot trial. Stroke survivors with upper-limb impairment are randomized to graded or fixed home training. Recruitment is stratified so that people without a smartphone are included, and they receive a loaner wrist sensor that needs no app interaction. Remote proxies are the weekly seven-motion FMA-UE estimate, daily primitive counts, and caregiver-reported Barthel scores.
At 3 and 6 months, a stratified random subsample receives a home visit. The visit includes a therapist FMA by an assessor blind to arm, and a blood draw for serum BDNF, S100B and myelin basic protein following the in-clinic study. Comparing caregiver Barthel reports against blinded FMA within the anchored subsample, separately by arm, estimates arm-differential report inflation directly. The blood markers serve as a secondary, report-independent anchor. We treat serum BDNF as noisy and do not make it primary.
Baselines
The naive reference is complete-case analysis: the arm difference in remote FMA-UE estimates among participants who stayed and wore the device. This is what a typical remote trial would report. Every other estimator must show smaller absolute bias than this in simulation and should move the pilot estimate in the direction predicted by the measured error profile.
The strongest existing approach to selection is model-based correction. One version is inverse probability weighting on observed baseline covariates, including baseline impairment, age, device ownership and education. The other is a two-stage Heckman selection model, as used for attrition in Life under stay-at-home orders. Both will use the same covariates, data and tuning budget as the anchored estimator, and the Heckman model gets its best-performing exclusion restriction. For proximal within-person effects we also fit weighted and centered least squares, the estimator from Assessing Real-Time Moderation. This is mainly to show that it targets a different question rather than to compete with our estimator.
The strongest existing measurement baseline is the remote proxy itself taken at face value. The best reported figure comes from Estimation of Fugl–Meyer Assessment Upper-Extremity Sub-Scores: total FMA-UE score R² = 0.9514, NRMSE = 0.0678, under leave-one-subject-out validation. For primitive identification, StrokeRehab reports a sequence-to-sequence action error rate of 0.30–0.34 on its stroke test set, against 0.39–0.45 for segmentation baselines. Beating these does not mean higher R². It means showing that a pooled R² near 0.95 hides severity-dependent bias large enough to shift a between-arm effect. If no such dependence appears, that baseline wins and we say so. No reported result was found on Lacuna for remote-proxy accuracy under unsupervised home use.
Two ablations isolate the key ingredient. The first draws the anchored subsample only from completers and device users, which removes the deliberate reach into the excluded. The second drops the report-versus-blinded comparison, which removes the self-report correction. If the full design does no better than the first ablation, the value lies in measurement calibration alone, not in sampling the lost.
Why not just…?
Why not just reweight with inverse probability weights or a Heckman model? If observed baseline severity fully explains who drops out and who is misscored, weighting is enough and much cheaper. That is a real threat. The concern comes from Partnering with a global platform and the older-adult think-aloud study, which show people leaving digital systems for reasons no covariate records. The pilot settles it empirically: if the weighted and anchored estimates agree within sampling error, we recommend weighting.
Why not just assess everyone in the clinic? Clinic-only follow-up has its own selection problem, because severely hemiplegic patients are the least able to travel. It also misses the daily dose information needed to test graded intensity. Even so, universal in-person follow-up is the unbiased gold standard. Our simulation reports bias and variance per assessor visit so that the cost saving of anchoring is explicit rather than assumed.
Why not just use passive sensors and drop self-report? This removes reporting inflation but keeps the measurement-level selection, since StrokeRehab already reports degradation for severe impairment. The stage-one audit measures that degradation directly. The anchored subsample then shows whether a sensor-only effect estimate is biased.
Why not just hand out loaner devices and simplify the interface? We do this, and it fixes ownership. It does not fix attrition or scoring error. Comparing loaner and owner strata within the pilot shows how much bias remains after access is equalized.
Proposed Experiments and Decisive Checks
The first decisive check is the stage-one audit. Does remote FMA-UE and primitive-count error grow with baseline impairment on StrokeRehab and the FMA-UE dataset? If the error is flat, the measurement-selection half of the claim fails.
The second check is the simulation, run with literature-derived attrition and the audited error profile. Across realistic settings, does complete-case analysis distort the graded-versus-fixed effect by more than the effect itself? Does weighting leave substantial bias that anchoring removes, at a subsample cost we report per percentage point of bias removed?
The third check is the pilot. Do the anchored and weighted estimates diverge? Does caregiver-reported Barthel inflation differ between arms? Does the blinded FMA result agree in direction with serum BDNF, as the animal work predicts?
A null result on all three would be informative too. It would mean remote rehabilitation trials can rely on weighting and sensor proxies, which is worth knowing before larger trials are designed.
References
- Gradually Increased Training Intensity Benefits Rehabilitation Outcome after Stroke by BDNF Upregulation and Stress Suppression
- Cognitive intervention combined with aerobic limb rehabilitation exercise promotes nerve and limb function recovery in hemiplegic patients with hypertensive intracerebral hemorrhage
- A remote measurement study of PTSD and cannabis use among veterans: Recruitment, retention, and data availability
- Impact of Push Notifications on Physical Activity and Sodium Intake Among Patients with Hypertension: Microrandomized Trial of a Just-in-Time Adaptive Intervention
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- Understanding How Older Adults Comprehend COVID-19 Interactive Visualizations via Think-Aloud Protocol
- StrokeRehab: A Benchmark Dataset for Sub-second Action Identification
- Estimation of Fugl–Meyer Assessment Upper-Extremity Sub-Scores Using a Mixup-Augmented LSTM Autoencoder and Wearable Sensor Data
- Exploratory Hidden Markov Factor Models for Longitudinal Mobile Health Data: Application to Adverse Posttraumatic Neuropsychiatric Sequelae
- Life under stay-at-home orders: a panel study of change in social interaction and emotional wellbeing among older Americans during COVID-19 pandemic
- Partnering with a global platform to inform research and public policy making