Quantifying Innovation in Stroke: Large Language Model Bibliometric Analysis.

Area of Science:

• Biomedical Engineering
• Health Informatics
• Neurology

Background:

• Stroke treatment innovations like thrombolysis and mechanical thrombectomy have advanced significantly over 30 years.
• Accurate quantification of stroke innovation is crucial for research and funding but lacks validated methods.
• Existing methods for measuring healthcare innovation are insufficient for identifying productive research avenues.

Purpose of the Study:

• To evaluate the correlation between stroke-related patents and scientific publications.
• To demonstrate the feasibility of using large language models (LLMs) for innovation analysis.
• To identify rapidly advancing stroke care innovations using an innovation life cycle model.

Main Methods:

• Bibliometric analysis of patents (European Patent Office) and publications (PubMed) from 1993-2023 using keywords 'stroke OR cerebrovascular'.
• A 13 billion-parameter Llama LLM was trained and validated (99.2% accuracy) to filter stroke-specific patents.
• Innovation clusters were identified, and patent/publication growth curves were analyzed to estimate life cycle stages.

Main Results:

• LLM filtering identified 28,225 relevant stroke patents from an initial 237,035.
• Seven innovation clusters emerged: pharmacological treatment, alternative medicine, rehabilitation devices, medical imaging, diagnostic testing, surgical devices, and AI methods.
• Pharmacological treatments show plateaued patent activity, while AI methods, rehabilitation devices, and medical imaging demonstrate exponential patent growth.

Conclusions:

• LLM application offers a scalable method for quantifying stroke innovation using patent and publication data.
• Pharmacological treatments appear to be in a saturation phase.
• AI methods, rehabilitation devices, and medical imaging represent high-growth areas for future stroke research and investment.