Introduction Large vessel occlusion (LVO) strokes account for the most severe cases of ischemic stroke and benefit significantly from early mechanical thrombectomy. Rapid and accurate diagnosis is crucial. Artificial Intelligence has emerged as a tool to enhance LVO detection on non‐contrast CT and CTA, yet reported diagnostic performance and clinical readiness vary widely. This SR&MA evaluate the pooled diagnostic accuracy of AI algorithms for LVO detection and assess factors affecting model performance and clinical translation. Methods Following PRISMA‐DTA guidelines, we searched MEDLINE, Embase, IEEE Xplore, and Scopus through May 2025. Studies were eligible if they evaluated AI tools for LVO detection in adult stroke patients using CT‐based imaging. Two reviewers independently screened studies and extracted data on study design, population, algorithm type, imaging modality, and performance metrics (sensitivity, specificity, AUC). A bivariate random‐effects model was used to pool diagnostic accuracy. Subgroup analysis compared classical ML (SVM, XGBoost), deep learning, and hybrid models. Risk of bias was assessed using the QUADAS‐2 tool, and publication bias was evaluated via Deeks’ funnel plot. Data synthesis and statistical analysis were conducted using R and Python. Results From 3128 screened records, 27 studies met inclusion criteria, comprising 23,987 scans across 11 countries. Overall pooled sensitivity and specificity for AI detection of LVO were 0.91 (95% CI: 0.88‐0.94) and 0.89 (95% CI: 0.85‐0.92), respectively. Deep learning models outperformed classical ML approaches, with convolutional neural networks (CNNs) achieving sensitivity of 0.91 and specificity of 0.90. XGBoost and SVM‐based models showed more variability across settings (sensitivities 0.87‐0.92), largely influenced by training dataset size and heterogeneity. Hybrid architectures (CNN with rule‐based postprocessing) demonstrated the highest pooled AUC (0.94). Only 6 studies reported external validation, and just 2 were prospective. Studies using CTA imaging had slightly better accuracy compared to non‐contrast CT alone. Common limitations included data imbalance, lack of blinding, and insufficient documentation of real‐time workflow integration. Conclusion AI models show high diagnostic accuracy for LVO detection, with deep learning and hybrid algorithms offering the most robust performance. However, the transition from algorithmic promise to clinical impact is constrained by limited validation, heterogeneity in methodology, and lack of deployment in real‐world settings. Our findings underscore the need for standardized benchmarks, transparent reporting, and prospective trials with diverse populations. Integration of AI into stroke pathways should focus not only on technical accuracy but also on usability, decision support alignment, and interpretability to ensure safe and effective implementation in time‐critical scenarios. image
Khabiry et al. (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: