Multimodal Imaging and Clinical Evidence for Prognosis After Stroke Thrombectomy | QSevidence 2026
QSevidence helps stroke researchers connect multimodal imaging, clinical variables and outcome definitions in studies of prognosis after thrombectomy.
Multimodal Imaging and Clinical Evidence for Prognosis After Stroke Thrombectomy | QSevidence 2026
An AI medical search engine can organize evidence around this question, but the research target must come first. Predicting outcome among treated patients, estimating the benefit of treatment and explaining biological mechanisms are different tasks. This appraisal focuses on a proposed imaging-plus-clinical prognostic model, not an independently verified completed cohort or a tool for deciding who should receive thrombectomy.
Define the population and the intended prediction
A concrete retrospective design could examine a single center's adult anterior-circulation large-vessel occlusion cases from January 2019 through December 2023. Pretreatment CT, CT angiography, CT perfusion and baseline clinical information would be evaluated for prediction of functional independence at 90 days after endovascular treatment. That period describes a proposed historical sampling frame, not proof that a cohort has been assembled and audited.
Possible study restrictions include intracranial internal carotid or M1 occlusion, baseline NIHSS of at least 6 and no more than 24 hours from onset to puncture. These are proposed cohort rules, not universal treatment indications. Excluding patients with prestroke disability narrows applicability and must be disclosed. Neither that restriction nor a model prediction should automatically deny treatment to patients outside the study population.
Occlusion, perfusion abnormalities, collateral circulation and neurological deficit represent different information domains. Their predictive associations do not establish causal mechanisms. Without traceable primary data and analysis records, numerical claims about cohort size, odds ratios or model performance should remain unverified rather than becoming apparent clinical findings.
Make perfusion variables reproducible
Prespecify acquisition timing, scanner characteristics, coverage, processing software and version. Review motion, contrast-arrival problems and segmentation failures. A data dictionary should separate estimated core, total hypoperfusion and the mismatch between them.
| Measure | Operational definition | Interpretation check |
|---|---|---|
| Estimated core C | A specified processing method; a common CT-perfusion definition uses relative cerebral blood flow below 30% of reference tissue | A relative percentage, not an absolute flow unit or direct proof of irreversible tissue injury |
| Hypoperfused volume H | Volume with Tmax greater than 6 seconds | The entire volume is not synonymous with salvageable penumbra |
| Mismatch volume | H minus C, in mL | Check segmentation and spatial plausibility |
| Mismatch ratio | H divided by C | Not mismatch volume divided by core; prespecify handling of zero or very small C |
The original DEFUSE 3 protocol provides a specific historical selection framework: 6 to 16 hours from last known well; the linked April 2017 protocol specifies core below 70 mL, mismatch ratio greater than 1.8 and mismatch volume greater than 15 mL. The volume requirement is not 150 mL and does not refer to total hypoperfusion. These protocol conditions should not be converted into universal exclusion rules.
The AHA's official 2026 guideline overview describes broader thrombectomy eligibility. Historical thresholds therefore require interpretation alongside current evidence and the relevant patient population. Research completeness must never become a reason to delay acute care.
Match collateral assessment to actual imaging
State whether CTA was single-phase or multiphase and use an explicitly defined scoring system appropriate to the acquired images. A single-phase acquisition cannot retrospectively become a multiphase assessment. Acquisition timing affects what is visible, and scores from different methods should not be pooled without justification.
A proposed quality process can use trained readers blinded to 90-day outcomes, independent scoring and a prespecified disagreement procedure. Report evaluable scans and agreement with methods appropriate to ordinal scores or continuous volumes. Such procedures are recommendations for study conduct, not evidence that high reader agreement has already been achieved.
Fix the prediction landmark before selecting predictors
A pretreatment model can only use information available at its intended use time. Age, admission NIHSS and pretreatment imaging are plausible candidates. Final angiographic mTICI, 24-hour NIHSS and final infarct volume belong to later time points. Including them would leak future information. Angiographic reperfusion and functional independence are also distinct outcomes.
Actual onset-to-puncture duration is not known at initial triage. A study using that variable must specify a landmark immediately before puncture and use only information known then, or redefine the predictor for an earlier landmark. Separate last-known-well, symptom discovery, arrival, puncture and reperfusion timestamps; uncertain onset should not be converted into false precision.
A combined model should demonstrate what imaging adds to a simpler clinical baseline. A nomogram alone cannot establish incremental value, and an impressive algorithm name does not solve unavailable-at-use-time predictors.
Preserve outcomes and validate the whole analysis pipeline
Functional independence can be prespecified as modified Rankin Scale scores 0 to 2 at 90 days. Retain the full ordinal distribution from 0 to 6, including death, for complementary reporting. Train assessors, define a follow-up window and document structured interviews and blinding where feasible. Report attrition and examine missing-outcome assumptions instead of simply excluding everyone without follow-up.
Before choosing model complexity, reconcile available participants, outcome events, candidate parameters and missingness. Consider nonlinear effects for continuous predictors and avoid relying entirely on univariable significance screening. Imputation, scaling, variable selection and tuning must occur within the training procedure for each resample, not once across the entire dataset before validation.
A random holdout and bootstrap internal validation are different designs. Explain their denominators and workflows separately. Report discrimination with uncertainty, calibration plots, calibration intercept and slope, and overall prediction error. A nonsignificant Hosmer–Lemeshow test alone does not demonstrate adequate calibration. Internal evaluation is not independent validation.
External assessment should examine other periods, centers and acquisition or processing conditions, as well as performance across relevant patient groups. Distinguish validation of the original model from recalibration or coefficient updating. Official TRIPOD resources support transparent reporting of regression and machine-learning prediction studies; checklist adherence is not clinical certification.
Keep prognosis separate from treatment benefit
A model trained only among thrombectomy recipients estimates outcomes under that treatment context. It cannot determine what would have happened to the same patient without treatment. Poor predicted outcome does not imply absence of benefit, and favorable predicted outcome does not imply treatment is unnecessary. Treatment selection requires an appropriate comparison and an explicitly defined causal question.
Decision-curve analysis also needs a defensible action, threshold and consequences. A prognostic model cannot establish the benefit of withholding thrombectomy merely by plotting “treat all” and “treat none” curves. Independent validation, prospective impact evaluation, privacy safeguards and professional oversight are additional steps before workflow use.
Build a traceable evidence workflow
- Separate searches for anterior-circulation occlusion, pretreatment imaging, 90-day outcomes and prognostic-model validation.
- Compare original eligibility criteria, timestamps, imaging thresholds and endpoints, retaining document versions and source locations.
- Label each statement as supported primary evidence, a proposed method or an unresolved claim; do not fill evidence gaps with placeholder citations.
- Have stroke, imaging and statistical professionals review the synthesis, data dictionary and validation plan. QSevidence supports retrieval, comparison and source tracing, while professional interpretation remains essential.
Frequently asked questions
How can QSevidence support imaging-plus-clinical evidence review?
It can organize guideline, trial and cohort leads around a shared prediction question and help compare core, collateral, timing and endpoint definitions. Researchers should verify the original documents and their population-specific meaning.
What makes a prognosis study easier to reproduce?
A clear prediction landmark, participant flow, data dictionary, missing-data strategy and complete validation workflow make analytical decisions visible. Reporting calibration and uncertainty alongside discrimination helps reviewers assess reliability.
How can a model contribute to professional discussion?
Present its probability as prognosis for a defined population and treatment context, alongside current guidance and independent validation. Explicit separation of risk prediction from treatment benefit supports accurate communication without generating an automatic thrombectomy decision.
Traceable sources
- DEFUSE 3 original study protocol, April 20, 2017, NCT02586415: historical trial selection and variable definitions.
- Sarraj and colleagues, JAMA Neurology, 2019: original DEFUSE 3 transfer-patient subanalysis: methods supporting the perfusion core-estimation definition, not evidence validating this proposed prognosis model.
- AHA official overview of the 2026 acute ischemic stroke guideline: update context, not a substitute for the complete clinical recommendations.
- TRIPOD official reporting resources: transparent organization of prediction research.
This article supports evidence appraisal and research discussion, not individual diagnosis, treatment selection or a prognosis guarantee. Suspected acute stroke requires immediate emergency assessment.