Evidence-based clinical practice guideline from the Neurocritical Care Society, addressing prognostic predictors and multivariate prediction models for functional outcome in ICU-admitted AIS patients
Acute ischemic stroke (AIS) is a leading cause of death and disability worldwide. Approximately 10–20% of AIS patients require admission to an intensive care unit (ICU) for management of life-threatening complications.
To provide evidence-based recommendations on the reliability of individual predictors and multivariate prediction models for functional outcome at 3 months in critically ill adults with AIS — specifically for use when counseling patients and their surrogates.
Prior to this guideline, there was no standardized, GRADE-based framework for neuroprognostication in this specific population. Most existing evidence comes from general AIS cohorts and cannot be directly applied to the critically ill.
This guideline used a narrative systematic review with GRADE methodology — the gold standard for clinical practice guideline development. Given expected heterogeneity in prognosis literature, a narrative (rather than meta-analytic) approach was chosen.
Initial librarian search: 20 February 2019 (covering 1946–2019) across MEDLINE/PubMed, EMBASE, Web of Science, Cochrane. Updated searches: 1 August 2022 and 5 February 2024.
| Criterion | Reason |
|---|---|
| Sample size < 100 adult patients | Insufficient statistical power for reliable prediction |
| TIA and/or mild stroke only | Not representative of critically ill AIS population |
| Highly selected subgroups (e.g., periprocedural stroke) | Indirectness — limits generalizability |
| Predictors not evaluated in multivariate analysis | Inability to isolate independent effect |
| Genetic polymorphism predictors | Not clinically applicable at bedside |
| Prediction models not reporting discrimination | Cannot assess model performance |
Used for studies of individual prognostic variables
Used for studies of clinical prediction models
The guideline specifically assessed risk of bias from self-fulfilling prophecy — where knowledge of a poor prognostic predictor leads clinicians to withdraw life-sustaining treatment, thereby causing the predicted poor outcome. Three specific domains were evaluated: (1) treatment suspension policy, (2) clinician blinding to predictor, (3) systematic use of predictor for prognostication during the study period.
Topic experts and a patient/family representative rated outcomes on GRADE 1–9 scale. Outcomes with median rating 7–9 were considered "critical."
The guideline selected predictors based on clinical relevance and the presence of an appropriate body of literature. A total of 9 clinical variables and 4 clinical prediction models were systematically evaluated.
Acute Stroke Registry and Analysis of Lausanne
Dense Artery, mRS, Age, Glucose, Onset-to-Treatment, NIHSS
Ischemic Stroke Predictive Risk Score
Totaled Health Risks in Vascular Events
Indirectness was a pervasive issue. Most studies of predictors following AIS were not limited to critically ill ICU patients. Therefore, the body of evidence was downgraded for indirectness across most PICOTS questions. Findings from general AIS populations may not fully apply to the sickest patients.
The following recommendations are primarily focused on prediction of functional outcome (mRS at 3 months). The guideline intentionally avoids recommendations for predicting mortality alone, due to high self-fulfilling prophecy bias in that literature.
Higher NIHSS scores are associated with worse functional outcomes. However, the quality of evidence is limited by indirectness (most studies from general AIS, not specifically ICU patients). Use as one component of a multimodal prediction approach, not in isolation.
Failure to achieve successful reperfusion (either via IV tPA, mechanical thrombectomy, or both) is associated with worse functional outcomes. Successful revascularization = stronger predictor when achieved.
ENI (typically defined as ≥4-point improvement on NIHSS at 24–72 hours, or reaching 0–1) is a robust predictor of good functional outcome. Absence of ENI portends poorer outcomes.
Better collateral circulation on imaging (CTA or conventional angiography) is associated with smaller infarct size and better outcomes. However, evidence quality is low due to variability in assessment methods and study populations.
Larger infarct volume on imaging is associated with worse outcomes, but studies varied widely in how and when infarct size was measured. No consensus on a specific threshold for prognostication in ICU patients.
Advanced age is associated with worse functional outcomes post-AIS. However, age alone should never be used as a sole predictor. Combine with other clinical and imaging variables for prognostication.
The following were rated as critical outcomes by experts but had insufficient evidence to support any recommendation:
• Blood glucose — hyper/hypoglycemia studied but conflicting data
• Hypertension history — inconsistent definitions and timing
• History of previous stroke — limited independent predictive value in multivariate models
The risk of self-fulfilling prophecy bias was deemed too high to recommend any single clinical variable for predicting death when "all available means of life support are used, indefinitely and without limitation." Supplementary data addresses mortality briefly but it was not a primary focus.
All four clinical prediction models were evaluated for discrimination (ability to separate those with good vs. poor outcomes) using standard metrics like C-statistic / AUC. However, all models showed significant limitations when applied to ICU-level critically ill patients.
None of the prediction models (ASTRAL, DRAGON, iScore, THRIVE) were developed or validated specifically in critically ill AIS patients. Their original development cohorts predominantly included patients from general stroke units and registries. When applied to ICU patients, calibration may be poor — predicted probabilities often overestimate or underestimate actual outcomes in this sicker subgroup.
| Model | AUC Range Reported | ICU Applicability | Key Limitation |
|---|---|---|---|
| ASTRAL | 0.75–0.85 | ⚠️ Moderate | Tested mainly in general AIS; lacks ICU-specific validation |
| DRAGON | 0.78–0.88 | ⚠️ Moderate | Includes pre-stroke mRS; performs well for 3-month outcome in general AIS |
| iScore | 0.72–0.82 | ⚠️ Low-Moderate | Includes comorbidities; developed from Canadian Stroke Registry |
| THRIVE | 0.68–0.80 | ⚠️ Low | Simplified model; limited discrimination in ICU populations |
These models can serve as adjunctive tools in the prognostic assessment of critically ill AIS patients, but should NOT replace clinical judgment or be used in isolation. They are best used within a multimodal, multidisciplinary prognostication framework that includes clinical examination, neuroimaging, neurophysiology (when available), and ongoing reassessment.
The majority of studies used the modified Rankin Scale (mRS), ranging from 0 (asymptomatic) to 6 (death). The guideline used an inclusive definition of good/poor outcome encompassing all mRS thresholds described in the literature — a deliberate choice given variability in how studies defined "good" vs. "poor" outcome.
Beyond specific variable recommendations, the guideline identified several GRADE "Good Practice Statements" — principles so fundamental they don't require formal evidence grading but represent current standard of care:
Neuroprognostication discussions should be honest, transparent, and calibrated. Avoid false optimism or unwarranted pessimism. Use probabilistic language ("chance of recovery is approximately X%") rather than deterministic statements.
Prognostic information should be presented to patients (if capable) and surrogates within a shared decision-making framework. Goals of care should be revisited repeatedly over time — prognosis is not static.
Neurological prognosis in AIS evolves over days to weeks. Single-timepoint predictions are insufficient. Serial clinical examinations, repeat neuroimaging, and neurophysiological studies (when indicated) should inform ongoing prognostic assessment.
Neuroprognostication should involve multiple specialties: neurology, neurocritical care, neurosurgery, rehabilitation medicine, nursing, and palliative care when appropriate. No single variable or model should drive decisions in isolation.
| Limitation | Impact |
|---|---|
| Indirectness — most evidence from general AIS, not ICU-specific populations | Recommendations may overestimate or underestimate true predictive value in ICU patients |
| Variable outcome definitions across studies (mRS thresholds) | Limited ability to perform meta-analysis; heterogeneity in findings |
| Self-fulfilling prophecy in mortality literature | Unable to make reliable recommendations for mortality prediction alone |
| No ICU-specific prediction models available | Existing models have limited calibration in critically ill AIS population |
| Patient-centered outcomes (QoL, cognition, depression) — insufficient data | Recommendations limited to functional outcome (mRS) |
• ICU-specific prediction models — none currently exist with adequate ICU validation
• Prospective studies in AIS-ICU patients with blinding to predictors to reduce self-fulfilling prophecy
• Patient-centered outcomes research — QoL, cognitive function, depression post-ICU in AIS survivors
• Standardized timing of prognostic assessment in the ICU course
• Biomarker studies (serum, CSF, neuroimaging) specifically in critically ill AIS cohorts