Abstract 053: Artificial Intelligence in Stroke Care: Bridging the Gap Between Innovation and Clinical Implementation

Introduction Stroke remains a global health challenge with high mortality and long‐term disability. Artificial Intelligence (AI), particularly machine learning (ML) and deep learning (DL), has emerged as a transformative tool across the stroke care continuum, from early detection to prognosis and rehabilitation. Despite promising results in research settings, integration into routine clinical workflows remains limited due to systemic, ethical, and validation‐related barriers. This abstract synthesizes performance metrics and translational barriers using a systematic approach to define the true clinical value of AI in stroke care. Methods We conducted a comprehensive meta‐synthesis of 33 studies evaluating AI applications in stroke care, including detection (MRI/CT), outcome prediction, and post‐stroke rehabilitation. Using the MI‐CLAIM checklist, we assessed model reporting fidelity, while the QUADAS‐2 tool evaluated risk of bias. We categorized applications into five domains: diagnosis, imaging, prognosis, rehabilitation, and implementation. Using R and Python, we performed pooled analysis to compute sensitivity, specificity, and AUC (Area Under the Curve), with subgroup comparisons across algorithm types (CNN, RF, SVM, XGBoost). One visualization was generated to summarize AI performance across domains. Results AI models demonstrated high pooled diagnostic performance for ischemic stroke on MRI with a sensitivity and specificity of 93%, validated by hierarchical summary ROC (HSROC) analysis. Prognostic models achieved an average AUC of 0.872, with SVM and XGBoost algorithms exceeding 0.90 in predictive accuracy. However, rehabilitation AI tools had heterogeneous results, with usability and customization being major determinants of success. Only 1 of 33 studies evaluated a CE‐marked AI tool, and just 15 met low‐bias thresholds. Major limitations include poor generalizability, lack of clinical trial validation, and algorithmic opacity. Importantly, stakeholder interviews revealed a disconnect between perceived AI utility and clinical trust, particularly in autonomous decision‐making roles. Conclusion While AI models exhibit strong technical performance in stroke diagnosis and prognosis, their real‐world implementation is hindered by clinical mistrust, under‐validation, and ethical concerns. Bridging this gap requires harmonized regulatory frameworks, rigorous clinical validation, and emphasis on human‐centered design. AI in stroke care is not yet a panacea, but a powerful adjunct, its future depends on inclusive development, transparent validation, and interdisciplinary cooperation. image