Systematic Review of Large Language Models and Natural Language Processing in Stroke Care: Applications, Challenges, and Future Directions

Stroke, a leading cause of mortality, manifests as ischemic (87%) or hemorrhagic (13%), demanding rapid intervention to mitigate irreversible damage. Despite advances in artificial intelligence, systematic reviews addressing the integration of large language models and natural language processing into clinical stroke care remain limited. Such a review is critical given large language models’ potential to overcome traditional natural language processing limitations, thereby enhancing risk prediction and decision support. We proposed a systematic review aimed to comprehensively analyze applications in stroke care. After searching 6 databases, 2991 records were screened, yielding 65 eligible studies. Results were structured quantitatively (impact factor trends, publication distribution) and qualitatively across 4 dimensions: Study Purposes (eg, risk modeling, decision support), Data Sets/Key Findings, Limitations, and Future Directions. Large language models demonstrated strong capabilities in automated data extraction from clinical notes (accuracies of 93.5%–95.1%) and report summarization. However, majority of studies (94%) lacked external validation. Most were limited by single-center, retrospective designs (62%) and used private data sets (85%), raising concerns about generalizability. Common failure modes included model hallucinations, performance degradation on external data, and infrastructural barriers to clinical integration. Future efforts must prioritize multicenter, prospective validation (82% of studies) to ensure model robustness and generalizability across diverse populations. Pathways for clinical translation include developing interpretability techniques to build clinician trust. Technical refinements for hallucination mitigation (98% of studies) and real-time integration of multimodal data are necessary to enhance predictive power. Addressing data heterogeneity and ethical concerns remains as gaps. This review highlights the potential of large language models in stroke care, encompassing tasks from risk prediction to workflow automation. Realizing this potential requires a shift from proof-of-concept studies to rigorously validated, clinically integrated systems. The field demands scalable, equitable, and transparent artificial intelligence solutions that are codeveloped with clinicians. These are needed to overcome existing methodological and translational barriers.