Neurology · Medical Literature
Predictive Performance of Artificial Intelligence Models for Stroke Risk Stratification: A Systematic Review and Meta-Analysis.
Abstract
BackgroundStroke remains a leading cause of mortality, long-term disability, and healthcare expenditure worldwide, placing substantial strain on healthcare systems, particularly in low- and middle-income countries. Effective risk stratification can facilitate targeted prevention strategies, optimize resource allocation, and reduce avoidable hospitalizations. This study synthesizes existing evidence on the predictive performance of artificial intelligence (AI)-based models for stroke risk assessment through meta-analysis and explores their potential implications for healthcare system planning.MethodsStudies were systematically retrieved from Web of Science (WoS), PubMed, and Scopus until 31 January 2025. The review followed the PRISMA 2020 guidelines. Area Under the Receiver Operating Characteristic Curve (AUC) values were extracted for each algorithm type and pooled using meta-analytic methods.ResultsDeep learning (DL) algorithms demonstrated favorable pooled discriminative performance (AUC: 0.955; 95% CI: 0.906-1.00, I2 = 85.75%), especially for imaging-based models. Sensitivity analysis modestly reduced heterogeneity (I2 from 85.75% to 61.77%). Substantial heterogeneity remained across study populations, healthcare settings, predictor characteristics, and validation strategies, limiting the generalizability of findings.ConclusionsAI-based models, particularly DL approaches, demonstrate favorable predictive performance for stroke risk stratification. However, considerable methodological heterogeneity, limited external validation, and risk of bias reduce confidence in widespread clinical implementation. Future research should follow standardized reporting and validation frameworks, such as TRIPOD, to improve methodological rigor, transparency, and clinical applicability.
Clinical relevance
Stroke remains a leading cause of mortality, disability, and healthcare expenditure globally, with disproportionate burden in low- and middle-income countries where resource allocation is critical. Current risk stratification tools have well-documented limitations, and AI-based models represent a rapidly emerging alternative whose aggregate predictive performance had not been formally synthesized until this meta-analysis. Neurologists and health systems planners need rigorous, pooled evidence to evaluate whether AI tools are ready for integration into stroke prevention pathways.
Key findings
1. Deep learning (DL) algorithms achieved the highest pooled discriminative performance for stroke risk stratification, with a pooled AUC of 0.955 (95% CI: 0.906–1.00), particularly strong for imaging-based models. 2. Substantial statistical heterogeneity was identified across DL studies (I² = 85.75%), indicating considerable variability in model performance attributable to differences in study populations, healthcare settings, predictor sets, and validation strategies. 3. Sensitivity analysis reduced but did not eliminate heterogeneity (I² from 85.75% to 61.77%), suggesting that outlying studies partially drive variability but that underlying heterogeneity remains clinically meaningful. 4. The systematic review followed PRISMA 2020 guidelines and drew on three major databases (PubMed, Web of Science, Scopus) with a search cutoff of 31 January 2025, providing a contemporaneous synthesis of the evidence base. 5. Limited external validation across included studies was identified as a key methodological limitation, reducing confidence in the generalizability of reported AUC values to real-world clinical deployment. 6. Risk of bias concerns were explicitly noted across studies, further constraining the strength of conclusions and precluding confident endorsement of widespread clinical implementation at this time.
Practice implications
Clinicians and health system leaders should recognize that while DL-based stroke risk models demonstrate impressive discrimination in research settings (pooled AUC ~0.955), the high heterogeneity and limited external validation mean these tools should not yet replace validated conventional risk scores (e.g., CHA₂DS₂-VASc, Framingham Stroke Risk Score) without institution-specific validation. Future procurement or development of AI stroke risk tools should mandate adherence to the TRIPOD reporting and validation framework, with prospective external validation in diverse populations—including those from low- and middle-income countries—before clinical adoption. Neurologists involved in quality improvement or digital health initiatives should advocate for standardized outcome definitions, transparent model reporting, and head-to-head comparisons with existing risk stratification instruments to establish incremental clinical value.