Artificial intelligence models in cerebral infarction: performance and applications in diagnosis, classification,...

Artificial intelligence (AI) holds significant promise for transforming cerebral infarction care, yet its real-world performance across the entire disease management continuum remains inadequately synthesized, with heterogeneous evidence and unclear clinical translation pathways. To address this gap, this systematic review and meta-analysis comprehensively evaluates the performance and application value of AI models in diagnosis, subtype classification, severity grading, treatment guidance, and prognosis prediction for cerebral infarction. Adhering to PRISMA 2020 guidelines, we identified 76 eligible studies (n = 45,289 patients) published between 2015 and 2025. Pooled estimates were derived using bivariate random-effects meta-analysis for diagnostic accuracy and random-effects models with restricted maximum-likelihood estimation for other outcomes. AI models demonstrated robust diagnostic accuracy (pooled sensitivity: 0.92, specificity: 0.89, AUC: 0.95), and consistently high performance in classification (accuracy: 0.87), grading (AUC: 0.91), treatment response prediction (AUC: 0.88), and 3-month functional outcome prediction (AUC: 0.90). Substantial heterogeneity was observed across diagnostic studies (I² = 68%), which was explored through pre-specified subgroup analyses. Risk of bias was low in 18 of 32 diagnostic accuracy studies per QUADAS-2, while prognostic studies had a mean Newcastle-Ottawa Scale score of 6.8, with common limitations in confounder control and follow-up completeness. Multimodal models integrating imaging and clinical data generally outperformed unimodal approaches. This synthesis provides robust, quantitative evidence supporting the integration of AI into clinical workflows to enable precision stroke care, while also delineating critical challenges-including dataset bias, limited interpretability, and infrastructural disparities-that must be addressed to facilitate successful clinical implementation.