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

Area of Science:

• Artificial Intelligence in Medicine
• Neurology
• Health Informatics

Background:

• Stroke is a leading cause of mortality, necessitating rapid intervention.
• Current artificial intelligence (AI) research in stroke care, particularly using large language models (LLMs) and natural language processing (NLP), requires comprehensive review.
• LLMs offer potential to enhance AI capabilities in stroke care beyond traditional NLP limitations.

Purpose of the Study:

• To systematically review and analyze the applications of LLMs and NLP in clinical stroke care.
• To identify key findings, limitations, and future directions for AI in stroke management.

Main Methods:

• Systematic review of 6 databases, screening 2991 records to identify 65 eligible studies.
• Quantitative analysis of publication trends and qualitative analysis across study purposes, data, findings, limitations, and future directions.

Main Results:

• LLMs demonstrate high accuracy (93.5%-95.1%) in automated data extraction and report summarization.
• Most reviewed studies (94%) lack external validation, and 62% use retrospective, single-center designs with private data (85%), limiting generalizability.
• Common issues include model hallucinations, performance degradation, and integration barriers.

Conclusions:

• LLMs show significant potential for stroke care tasks, including risk prediction and workflow automation.
• Further research must focus on multicenter, prospective validation and address generalizability, interpretability, hallucination mitigation, and ethical concerns.
• Clinical translation requires validated, integrated AI systems co-developed with clinicians to overcome existing barriers.