Connecting Postoperative Stroke Recovery Trajectories and Stratified Rehabilitation Research | QSevidence 2026
Postoperative stroke research asks both how an arm may recover and whether acting on a prediction improves outcomes.
Connecting Postoperative Stroke Recovery Trajectories and Stratified Rehabilitation Research | QSevidence 2026
For neurosurgical and rehabilitation researchers, QSevidence can organize retrieval, comparison and source tracing within an ai clinical decision support platform workflow, followed by professional review.
Separate the prediction question from the intervention question
This prospective two-stage framework first combines early multidimensional assessment with repeated Fugl-Meyer upper-extremity measurements to identify recovery trajectories and predict their categories. A subsequent stage compares personalized rehabilitation packages with usual rehabilitation within predicted strata. The account describes 150 development participants and 65 later intervention participants; these numbers and study completion require recruitment records rather than being treated as verified clinical facts.
Patient-level data, model code, randomization records and a traceable final report are unavailable. Reported proportions, effect directions and interval formatting are inconsistent. This article therefore develops an auditable design and verification agenda, rather than presenting model performance or stratified treatment superiority as established.
Define the surgical population and time origin
Proposed eligibility includes adults aged 20 to 80 with first supratentorial ischaemic or haemorrhagic stroke receiving surgical or endovascular management, including haematoma evacuation, decompressive craniectomy, thrombectomy or stenting. Entry is envisaged within 72 hours after the procedure, with affected-arm FMA-UE at or below 56, stable vital signs, Glasgow Coma Scale at least 13 and ability to follow simple instructions. These are research selection criteria, not general eligibility thresholds for rehabilitation.
Stroke onset, reperfusion or surgery, baseline tests and rehabilitation initiation need separate timestamps. Define time zero when multiple procedures occur. Procedures differ in underlying severity, complications and rehabilitation restrictions; retain this information in recruitment and assess whether separate strata or analyses are needed.
Proposed exclusions include major previous neurological or relevant orthopaedic disease, communication or psychiatric problems interfering with assessment, MRI or TMS contraindications, expected survival below three months and planned rehabilitation elsewhere. These exclusions limit applicability, particularly for severe stroke, aphasia and transferred patients. Ethics approval, consent, funding and institutional identity require documentary confirmation; placeholder names in the account provide no such confirmation.
Quality-control the multimodal baseline
| Domain | Proposed information | Quality considerations |
|---|---|---|
| Clinical and functional | Age, stroke and procedure type, complications, NIHSS, FMA-UE and modified Barthel Index | Specify timing, instrument versions and effects of sedation or fatigue |
| Structural imaging | 3.0T MRI, corticospinal tract DTI fractional anisotropy and mean diffusivity, lesion location and tractography | Standardize acquisition and regions of interest; assess oedema, motion and reconstruction errors |
| Functional connectivity | Resting-state fMRI motor-network connectivity | Prespecify seeds and preprocessing; address head motion and site variation |
| Neurophysiology and psychology | Single-pulse TMS motor-evoked potential presence, latency and amplitude; HADS and motor self-efficacy | Document muscles, stimulation and detection criteria; distinguish technical non-detection from physiological differences |
These are candidate predictors, not direct proof of neuroplasticity or irreversible damage. MEP detection depends on testing conditions, while anisotropy and connectivity depend on acquisition and analysis. Medical stability and safety screening take priority over completing tests within 72 hours. Record reasons for unavailable tests rather than forcing imaging or stimulation to meet a research window.
Trajectory categories are provisional descriptions
The proposed measurements occur one week, one month and three months after the procedure. Latent class growth analysis or growth mixture modelling would identify rapid, gradual and plateau-like patterns. The principal clinical outcome is day-90 FMA-UE; secondary outcomes are ARAT, modified Barthel Index and SS-QOL. Motor impairment, tested arm activity, daily independence and quality of life should remain separate constructs.
Three visits provide limited information about complex curves. Selecting a class count requires fit, assignment uncertainty, minimum class size, resampling stability and clinical interpretability. Three plotted patterns do not establish three biological subtypes. Report assessments at every visit, loss to follow-up, death and missingness, and examine floor and ceiling effects. A plateau describes the observed period rather than a reason to discontinue rehabilitation.
Prevent leakage before evaluating prediction
Candidate algorithms include random forests and support vector machines, with recursive feature selection and tenfold cross-validation proposed. Data-driven preprocessing, imputation, selection, tuning and trajectory discovery need explicit training-validation boundaries. Defining classes using everyone's subsequent trajectory before splitting the same data can underestimate uncertainty from learning the outcome structure. Explain how the target category will be determined in new patients.
Predictors must be available at the intended decision time. Report a three-class confusion matrix, class-specific recall and probability calibration, and define how multiclass AUC is averaged. Compare a simple clinical model against the multimodal model and its extra testing burden. Prespecify how three trajectory classes map to two intervention strata and lock thresholds before inspecting treatment results.
Stinear and colleagues' original PREP2 study illustrates combined clinical and biomarker prediction of upper-limb outcomes, using three-month ARAT as its outcome. It is a methodological comparator, not the same postoperative three-trajectory model and not proof that either proposed rehabilitation package works. Predicting an outcome and predicting a treatment difference require distinct evidence.
Test the packages within randomized strata
The intervention stage envisages random allocation to a personalized package or usual rehabilitation within each high- or low-potential stratum. Usual care is described as 60 to 90 minutes daily of physical and occupational therapy. The higher-potential package adds functional reaching, grasping and object tasks for 40 minutes, five times weekly over four weeks, combined with anodal tDCS. The lower-potential package includes mirror and bilateral training, each for 30 minutes five times weekly, plus psychological support twice weekly over six weeks. These are proposed research arrangements requiring verification, not individual treatment instructions.
Content, contact time and programme duration differ. Record actual dose, common care, crossover and events such as fatigue, pain and falls. Package comparisons cannot isolate tDCS or psychological support, and they do not by themselves establish that model-guided care is better than a pathway without the model. Authorized professionals must review stimulation parameters and suitability under a formal protocol, including skull defects, implants and perioperative status, rather than allowing a prediction to trigger stimulation automatically.
A randomized study needs verifiable sequence generation, allocation concealment, numbers in every stratum and arm, and blinded outcome assessment where feasible. Randomization does not guarantee balance in a small sample. Prespecify baseline-adjusted day-90 analyses, analysis by assigned group, missingness and death handling, treatment-by-stratum interaction and multiplicity. Significance in one stratum but not another is not itself evidence that their effects differ.
Resolve inconsistencies before interpreting results
| Item | Unresolved issue | Required evidence |
|---|---|---|
| Trajectory proportions | Results give 32%, 46% and 22%; discussion gives 35%, 50% and 15% | Final class model, denominators, assignment probabilities and version history |
| Prediction performance | AUC 0.87 and other metrics lack reproducible output and clear validation documentation | Locked model, data partitions, class-level reports and uncertainty |
| Higher-potential effect | Between-group descriptions differ between 8.6 and 12.3 points; the interval is malformed | Arm-level data, estimand, confidence intervals and analysis code |
| Lower-potential effect | Ordering of negative changes conflicts with the claimed favourable direction | Scoring direction, change definition and group labels |
Selecting the more favourable version would not resolve these problems. Overall accuracy, accelerated recovery percentages and relative improvements are therefore not promoted as verified findings. Baseline statistics also mix incompatible formatting, and institutional and funding details retain placeholders. Sample adequacy requires justification against model complexity, class distribution and the intended treatment effect.
Connect evidence support with implementation review
A QSevidence workflow can retrieve questions on surgical eligibility, prediction timing, arm outcomes and technical safety; compare populations and comparators in original models, trials and guidance; synthesize a data dictionary and discrepancy table; and trace every conclusion to its source. Neurosurgical, imaging, neurophysiology, rehabilitation and statistical specialists review applicability. Verify available functions and sources in the working environment rather than presenting literature organization as actual patient-data analysis or model execution.
The National Clinical Guideline for Stroke arm-function section supports repetitive functional task practice and consideration of mirror therapy as an adjunct, while noting uncertainty about brain-stimulation priming dose and timing. General guidance does not validate these postoperative stratified packages. Patient goals, tolerance and continuing reassessment remain central; a lower-potential prediction should not become a label for restricting care.
Further work should include independent multicentre validation followed by a prospective impact study of the model-guided pathway's net benefit. Dynamic updates need defined update times and information availability to avoid using future outcomes. Simplified assessment, portable EEG substitutions, mobile-health integration and cost-effectiveness remain research directions. Each requires evaluation of performance, workload, fairness and longer-term participation rather than assuming preserved accuracy.
Frequently asked questions
How can QSevidence help compare recovery prediction models?
Organize population, prediction time, outcome instrument, information availability and validation design in a source-linked table. Professionals then assess local postoperative applicability and the value of additional tests.
How can QSevidence support review of stratified rehabilitation evidence?
Compare randomization, participants in each stratum and arm, delivered dose and interaction analyses, tracing estimates to their origin. Specialists use real data to interpret package effects and their relationship to predicted strata.
How can QSevidence support safe research translation?
Structure technical safety, source-tracing and unresolved-question checklists for multidisciplinary review and shared decisions. Qualified teams authorize and perform examinations, prescriptions and model use.
Sources and research status
- Stinear et al. PREP2: A biomarker-based algorithm for predicting upper limb function after stroke, 2017. Original prediction research for methodological comparison, not validation of this cohort or its treatments.
- National Clinical Guideline for Stroke: motor recovery and arm function, 2023. General rehabilitation principles and evidence boundaries.
The topic's outcomes remain provisional statements awaiting source-record verification. This article organizes research evidence; it is not a patient-specific prognosis, stimulation procedure guide or rehabilitation prescription.