Large Language Models in Stroke Management: A Review of the Literature

Stroke care generates vast free-text records that slow chart review and hamper data reuse. Large language models (LLMs) have been trialed as a remedy in tasks ranging from imaging interpretation to outcome prediction. To assess current applications of LLMs in stroke management, we conducted a narrative review by searching PubMed and Google Scholar databases on January 30, 2025, using stroke- and LLM-related terms. This review included fifteen studies demonstrating that LLMs can: (i) extract key variables from thrombectomy reports with up to 94% accuracy, (ii) localize stroke lesions from case-report text with F1 scores of 0.74–0.85, and (iii) forecast functional outcome more accurately than legacy bedside scores in small pilot cohorts.

These results, however, rest on narrow, retrospective datasets-often from single centers or publicly available case reports that the models may have encountered during pre-training. Most evaluations use proprietary systems, limiting reproducibility and obscuring prompt design. None stratify performance by sex, language, or socioeconomic status, and few disclose safeguards against hallucination or data leakage.

We conclude that LLMs are credible research tools for text mining and hypothesis generation in stroke, but evidence for clinical deployment remains preliminary. Rigorous, multisite validation, open benchmarks, bias audits, and human-in-the-loop workflows are prerequisites before LLMs can reliably support time-critical decisions such as thrombolysis or thrombectomy triage.