Abstract

Rapid prehospital identification of stroke is essential to reduce time to definitive treatment, yet symptom-based screening tools used by emergency medical services (EMSs) have variable accuracy. This systematic review and meta-analysis evaluated the diagnostic accuracy of artificial intelligence (AI) models for prehospital stroke detection and summarized comparisons with commonly used prehospital scales when reported. We followed the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses guideline and prospectively registered the protocol in the International Prospective Register of Systematic Reviews (PROSPERO; No. CRD420251091197). In July 2025, we searched MEDLINE (via PubMed), EMBASE, the Cochrane Central Register of Controlled Trials, and Google Scholar and screened records using predefined eligibility criteria. We extracted diagnostic accuracy outcomes, assessed risk of bias using the Quality Assessment of Diagnostic Accuracy Studies-2 tool, and pooled sensitivity and specificity for ischemic and hemorrhagic stroke using random-effects models. Sixteen studies were included. Pooled sensitivity and specificity for ischemic stroke were 86.9% (95% confidence interval [CI]: 69.9%-95.0%) and 88.6% (95% CI: 77.8%-94.5%), respectively. For hemorrhagic stroke, pooled sensitivity was 90.6% (95% CI: 86.2%-93.6%), and specificity was 93.9% (95% CI: 87.6%-97.2%). Diagnostic performance varies across studies, consistent with heterogeneity in model types, input data, reference standards, and clinical settings. Overall, AI algorithms have shown promise regarding prehospital stroke identification, although significant heterogeneity as well as divergent performance make it hard to be certain. Prospective validation studies, among other investigations, are hence required to optimize their use as auxiliary diagnosing devices.

Key words: Stroke, emergency medical services, artificial intelligence, machine learning, diagnostic accuracy, sensitivity and specificity