After a stroke, many people lose the ability to move one finger without the others, or one hand muscle without its opponent. Physiologists now measure this loss directly, recording dozens of forearm electrodes and showing that separate muscle groups fuse into fewer, coarser units. Rehabilitation trials, however, still largely judge success by what patients say they can do. This proposal asks a question nobody in the supplied literature has tested: when patients rate their own hand function, can they perceive loss of finger selectivity at all? Or do they report whether a task got done, regardless of how it got done? If the second is true, patient questionnaires are not only noisy. They are blind to the very deficit that separates neurological recovery from learned compensation.
The study pairs trial-by-trial self-ratings with simultaneous finger kinematics and high-density electromyography (EMG, recordings of the electrical activity of muscles from a dense electrode grid). It then splits the gap between self-report and physiology into two parts. The first is an introspective part: patients cannot sense the deficit. The second is a social part: patients shade their answers for the person asking. The social part is estimated by randomizing who receives the rating. The result would tell rehabilitation researchers which self-report items can be trusted as proxies for selectivity, and which ones need a sensor.
Core Claim
We claim that self-report error about hand function in chronic stroke is structured, not random. It is driven mainly by an introspective blind spot for selectivity loss, and the blind spot is largest when the task still succeeds. Two cases fit this description. In the first, antagonist co-contraction stiffens the hand while it still moves: agonist and antagonist muscles fire together, so the joint locks up. In the second, grip force builds normally but release fails.
The claim predicts three measurable things. A patient's immediate, recall-free rating of "did only that finger move?" should poorly discriminate selective from non-selective trials, even though ratings of "did the task succeed?" discriminate success well. The discrimination failure should be worse for co-contraction-dominated deficits than for overt synergy merging. Randomizing the rating recipient (private tablet versus the therapist) should shift the average level of ratings without restoring the ability to tell good trials from bad ones.
This framing addresses a tension in the file. Two groups find that spatial synergy structure is relatively preserved in chronic stroke hands, with deficits driven by co-contraction and altered common drive: Identifying alterations in hand movement coordination from chronic stroke survivors using a wearable high-density EMG sleeve and Leveraging neural drive to assess hand motor function in individuals with chronic stroke. Reaching studies instead describe impairment as merging of healthy modules: Alterations of Muscle Synergies During Voluntary Arm Reaching Movement in Subacute Stroke Survivors at Different Levels of Impairment and Muscle synergy analysis yields an efficient and physiologically relevant method of assessing stroke. We do not try to settle which picture is correct. We use the fact that patients differ along this axis as a natural moderator of how visible the deficit is from the inside.
Relation to Prior Work
The closest work validates sensor features against clinician scores, not patient reports.
- Motivating work (the physiology exists, but self-report is unaddressed). Three studies map sensor features onto the upper-extremity Fugl-Meyer Assessment (FMA-UE), a clinician-rated impairment scale.
- The wearable EMG sleeve study combines co-contraction indices and matrix-factorized synergies into one principal component that tracks the FMA hand subscore (art_3b5a0f3c).
- The neural-drive study decomposes the same 75-channel recordings into single motor-unit firing and forms a comparable signature (art_ce1fbcf9).
- Estimation of Fugl–Meyer Assessment Upper-Extremity Sub-Scores Using a Mixup-Augmented LSTM Autoencoder and Wearable Sensor Data predicts subscores from four motion sensors.
- All three treat the clinician score as ground truth and never ask the patient anything. We keep their sensors but change the target: the patient's own judgment, measured against the sensors.
- Motivating work (the item-level pattern suggests where error concentrates). The Brain 2022 synergy study found that only 26 of 37 FMA items track synergy merging.
- Constraining work (where the blind spot is likely). A low-dimensional representation of arm movements and hand grip forces in post-stroke individuals shows that grip force release is specifically distorted while force buildup is preserved. This tells us to include a release task. It also warns that task-completion measures, and by extension "could you do it?" questions, will miss release deficits.
- Constraining work (low-cost vision will not see fingers). Automated Movement Assessment in Stroke Rehabilitation found that wrist tracking alone suffices for its therapist rubric. That rubric is about task segments, so this rules out camera-only wrist tracking as a selectivity reference.
- Self-report literature (mechanisms we must separate, with no physiological anchor).
- Social desirability and inflated self-assessment appear in Heng Yue et al. and the new-nurses survey.
- Demand characteristics in intervention trials appear in Leong Bin Abdullah et al. and Surana et al..
- State-dependent recall appears in Vernon et al..
- None of these can say how much of a biased score is honest misperception, because none has a physiological truth to compare against. Our sensors supply that truth for one narrow construct.
- Methods we reuse.
- The micro-randomized trial logic of Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns randomizes a condition at each decision point and estimates effects moderated by current state. We apply it to the rating recipient.
- Its "negative moderation" finding, that prompts hurt people who were already doing well, is a caution: the act of asking can change behavior. We measure this explicitly.
- The measurement-invariance testing in the cross-lagged panel study checks that a questionnaire means the same thing across groups or times. We use it to test whether rating items mean the same thing in private versus therapist-facing conditions.
Research Program
Cohort. We recruit chronic stroke survivors spanning mild to moderate hand impairment. The target is roughly 40, far larger than the 7-patient sleeve cohorts. We add age-matched controls, who give the calibration ceiling.
Recording. Each participant wears the 75-channel forearm EMG sleeve protocol from the two sleeve papers, plus finger-joint kinematics from a glove or multi-camera hand tracking. They perform the same 12 cued hand and wrist movements. We add two things:
- a single-finger individuation block, which moves one finger while measuring involuntary movement of the others, called "enslavement";
- a grasp-and-release block on an instrumented object that measures grip force buildup and release.
Sensor-derived truth per trial. Each trial gets three physiological measures:
- a kinematic individuation index: movement of the instructed finger relative to total finger movement;
- an EMG co-contraction index, following the sleeve study;
- the similarity of that trial's synergy weights to the participant-pooled healthy templates, following the reconstruction approach of Pan et al.
Force-release trials get a release-duration measure.
Trial-level self-report. Before each trial, the participant predicts on a slider whether only the instructed finger will move. Immediately after, they rate two separate things: whether the task succeeded, and whether only the instructed finger moved. Collecting ratings trial by trial removes recall from the picture.
The social manipulation. On each trial, the recipient of the post-trial rating is micro-randomized. Either it is entered privately on a tablet with an explicit statement that therapists never see it, or it is spoken to the therapist running the session.
Transfer to everyday self-report. Participants complete standard retrospective hand-function questionnaires at the end of the session and again after one week of daily smartphone ratings. This tests whether lab-measured introspective calibration predicts error in the instruments trials actually use.
Core model. The post-trial selectivity rating $r_{it}$ of participant $i$ on trial $t$ is regressed on the sensor individuation index $s_{it}$, the randomized recipient indicator $d_{it}$, and a participant-level deficit profile $m_i$. The profile $m_i$ is the balance of co-contraction versus synergy merging from the physiology.
$$r_{it} = \alpha_i + (\beta_0 + \beta_1 m_i)\, s_{it} + \gamma\, d_{it} + \epsilon_{it}$$
The terms are read as follows:
- The slope $\beta_0 + \beta_1 m_i$ is introspective sensitivity: how much the rating moves when true selectivity moves.
- $\gamma$ is the causal shift in ratings due to the social condition.
- $\alpha_i$ is a person-specific offset.
- Our hypotheses are three: sensitivity for selectivity is low relative to sensitivity for task success; $\beta_1$ shows that sensitivity falls as the deficit profile shifts toward co-contraction; and $\gamma$ is positive while the social condition leaves the slope unchanged.
We also report a discrimination score: the area under the ROC curve (AUROC) of each participant's rating in separating selective from non-selective trials, using a sensor threshold calibrated on controls.
Baselines
Naive reference. The do-nothing reference assumes self-report is unbiased: the retrospective questionnaire score is taken as the selectivity measure, with no sensor correction. Against this, the question is simply how far questionnaire scores sit from sensor truth and in which direction. Beating it means showing a systematic, deficit-dependent signed error, rather than error with a zero mean.
Clinician-score proxy. This baseline predicts each patient's self-report error from their FMA-UE hand subscore alone. It is the cheapest thing a clinic could do, and it is the target that all the existing sensor papers optimize. The strongest reported sensor-to-clinician link in the file is $R^2 = 0.86$ between the combined co-contraction and synergy principal component and the FMA hand subscore (art_3b5a0f3c; Identifying alterations in hand movement coordination from chronic stroke survivors using a wearable high-density EMG sleeve). Close behind are $R^2 = 0.81$ for the motor-unit signature (art_ce1fbcf9) and $R^2 = 0.8264$ for four-sensor prediction of hand subscore Part C (art_c87cbfa5). These are agreements with clinician ratings, not with patient reports.
For the actual target, predicting patient self-report error from physiology, there is no reported result found on Lacuna. Our sensor model beats this baseline if it explains substantially more of the variance in self-report error than the FMA subscore does, under leave-one-subject-out cross-validation with identical folds. If the FMA subscore alone explains most of the error, the sensors are not needed for this purpose, and we will say so.
Kinematics-only model. This is the strongest competitor and is described under the first Why-not question below. It uses the same regression and folds, but with finger kinematics only.
Kinematic model plus scalar EMG. An ablation adds a single scalar EMG co-contraction index to the kinematic model, without synergy decomposition. This isolates whether the full high-density decomposition earns its cost over a crude EMG summary. Every model gets the same features budget and the same nested tuning.
Why not just…?
Why not just measure finger kinematics and skip the EMG sleeve? This is a real threat. Enslavement is visible in joint angles, so a glove might predict self-report error as well as the full EMG pipeline. The case for EMG is narrow and testable. Co-contraction can stiffen a hand that still reaches the right posture, and compensatory recruitment can reproduce normal trajectories; both points appear in the sleeve paper and in Pan et al. Those are exactly the cases where we predict patients rate themselves well. The kinematics-only model is therefore run as a full competitor. If adding EMG does not improve prediction of self-report error, specifically in the co-contraction-dominated participants, the proposal's central mechanism fails. We would then recommend gloves.
Why not just compare questionnaires to the clinician's Fugl-Meyer score? It costs nothing, and the discrepancy already exists in every trial dataset. But the FMA hand subscore is itself a visual, coarse rating, as the sleeve and neural-drive papers note. The Brain 2022 study shows that a third of FMA items do not reflect synergy merging at all. A patient-versus-clinician gap therefore mixes two biased observers. The clinician-proxy baseline above measures exactly how much this cheaper approach recovers.
Why not just administer questionnaires more often, as daily smartphone ratings? Frequent ratings remove recall distortion. Our trial-level ratings are that idea taken to its limit: zero delay. The core prediction is that even immediate ratings discriminate selectivity poorly, because the problem is perception, not memory. If immediate selectivity ratings turn out to discriminate well and only retrospective questionnaires fail, then frequent smartphone sampling is the right fix. That outcome would redirect the field toward sampling instead of sensors. Frequent prompting also carries the negative-moderation risk documented by NeCamp et al., which we monitor through the daily-rating week.
Decisive Checks
The study stands or falls on five checks:
- the participant-level AUROC for selectivity ratings compared with task-success ratings;
- the sign and size of $\beta_1$, the dependence of introspective sensitivity on co-contraction versus merging;
- a slope comparison across social conditions, using the invariance testing of the cross-lagged panel study, to confirm that the social condition shifts level without restoring discrimination;
- whether lab-measured sensitivity predicts, out of sample, the signed error of the retrospective questionnaire against the sensor battery;
- whether force-release deficits are underreported relative to buildup deficits, a direct test of the release blind spot suggested by the grip-force study.
A null result on the first check, meaning patients discriminate selectivity well, would falsify the introspective-blindness claim outright. We treat that outcome as informative: it would mean selectivity-specific questionnaire items are usable as they stand.
References
- Identifying alterations in hand movement coordination from chronic stroke survivors using a wearable high-density EMG sleeve
- Leveraging neural drive to assess hand motor function in individuals with chronic stroke
- Alterations of Muscle Synergies During Voluntary Arm Reaching Movement in Subacute Stroke Survivors at Different Levels of Impairment
- Muscle synergy analysis yields an efficient and physiologically relevant method of assessing stroke
- Estimation of Fugl–Meyer Assessment Upper-Extremity Sub-Scores Using a Mixup-Augmented LSTM Autoencoder and Wearable Sensor Data
- A low-dimensional representation of arm movements and hand grip forces in post-stroke individuals
- Automated Movement Assessment in Stroke Rehabilitation
- Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders
- The effects of work readiness, organizational justice and professional identity on the work performance of new nurses: a cross-sectional survey
- Acceptance and commitment therapy versus mindfulness-based stress reduction for newly diagnosed head and neck cancer patients: A randomized controlled trial assessing efficacy for positive psychology, depression, anxiety, and quality of life
- Tag2Risk: Harnessing Social Music Tags for Characterizing Depression Risk
- Incidence and Prevalence of Post-COVID-19 Myalgic Encephalomyelitis: A Report from the Observational RECOVER-Adult Study
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- Longitudinal relationships among problematic mobile phone use, bedtime procrastination, sleep quality and depressive symptoms in Chinese college students: a cross-lagged panel analysis