Answers are grounded in this paper and cite their sources.
In the triage of acute ischemic stroke (AIS), rapid non-contrast computed tomography (NCCT), computed tomography angiography (CTA), and diffusion-weighted magnetic resonance imaging (DWI) guide intravenous thrombolysis and endovascular thrombectomy (EVT). Machine learning and deep learning models have reported high discrimination for ischemic lesion detection, malignant cerebral edema (MCE), and 90-day modified Rankin Scale (mRS) outcomes. However, opaque model architectures hinder clinical interpretation. Materials and Scope: This critical review is based on a structured PubMed/MEDLINE query (1 January 2021 to 15 August 2026). Among 34 candidate records, two were excluded during initial screening, and two were unretrievable, yielding 30 fully appraised primary prediction studies (24 acute-AIS core studies and six adjacent neurovascular models). Eight acute-AIS studies were prioritized for detailed main-text comparison, while the remaining 22 studies were comprehensively retained in supplementary synthesis. All 30 available models underwent dual evaluation: prediction-model risk of bias via PROBAST+AI and an independent eight-dimension XAI validity appraisal. Overall risk of bias was High in 93.3% (28/30; driven primarily by Domain 4, analysis) and Unclear in 6.7% (2/30; Li et al. and Huang H. et al.); no study was rated Low. Multicollinearity and Attribution Sensitivity: In radiomic and hemodynamic modeling, strong dependence among interdependent imaging predictors (e.g., ASPECTS, core infarct volume, Tmax > 6 s hypoperfusion volume, and white matter hyperintensities) complicates feature-level SHAP interpretation. Interventional or other marginal-reference formulations can break the dependence between observed and integrated-out features and may evaluate combinations that are poorly supported by the empirical joint distribution. Consequently, individual-feature attributions may be sensitive to the explainer formulation and background/reference distribution and should not be interpreted as unique biological or causal effects. Coarse Saliency and Modality Bias: Grad-CAM generates spatially coarse attribution maps due to convolutional downsampling, making it susceptible to skull-base CT beam-hardening artifacts and MRI susceptibility artifacts. Furthermore, unimodal Grad-CAM architectures inherently fail to contextualize non-imaging clinical covariates (e.g., baseline NIHSS, onset-to-door time). Validation-stage Discrimination and Calibration: Across multicenter studies, cross-cohort discrimination is heterogeneous rather than uniformly attenuated. Cao et al. showed marked internal-to-external attenuation (AUC 0.918 to 0.755), whereas Huang L. et al. (0.78 to 0.88) and Li et al. (0.84 to 0.86) showed preserved or improved external discrimination. Training/development, internal validation, and independent external-validation stages must be distinguished. Calibration and uncertainty quantification (UQ) require separate evaluation; explanation maps alone do not establish reliable clinical probabilities. XAI must transition from descriptive post hoc visualizations into validated clinical decision support. Safe translation requires explicit reporting and sensitivity analysis of feature-dependence handling, quantitative saliency validation (e.g., lesion-mask intersection-over-union [IoU] or Pointing Game hit rates against expert anatomical segmentations, alongside deletion/insertion fidelity curves), multi-vendor image harmonization, and uncertainty quantification. Grouped or hierarchical attribution may be considered when appropriate to the intended explanatory target.
No takes yet. Share an insight, caveat, or question.
Cheng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: