Semi-Automating Curation of Clinical Practice Guidelines.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Knowledge Management

Background:

• Clinical knowledge is expanding rapidly, making it difficult to keep clinical practice guidelines (CPGs) current.
• Manual curation of CPGs is time-consuming and leads to delays in incorporating new evidence.
• Existing methods struggle to efficiently integrate the latest research findings into clinical recommendations.

Purpose of the Study:

• To develop and validate a reproducible pipeline for the semi-automated surveillance of clinical practice guidelines (CPGs).
• To enable faster and more efficient updates to CPGs by leveraging artificial intelligence.
• To reduce the burden of manual curation in guideline maintenance.

Main Methods:

• Ingested heterogeneous guideline documents and parsed text using PyMuPDF and spaCy.
• Normalized terms through a concept layer and embedded passages using biomedical language models.
• Indexed passages with FAISS for dense retrieval and employed retrieval-augmented generation (RAG) for drafting suggestions.

Main Results:

• The developed pipeline successfully ingested, processed, and retrieved information from guideline documents.
• PubMedBERT demonstrated superior performance in embedding and retrieval tasks.
• The retrieval-augmented generation approach produced editorial suggestions with provenance, achieving 80% inter-annotator agreement.

Conclusions:

• A semi-automated pipeline can effectively support the surveillance and updating of clinical practice guidelines.
• AI, particularly using models like PubMedBERT within a RAG framework, can significantly enhance the efficiency of guideline maintenance.
• This approach facilitates expert-guided updates, ensuring CPGs remain current and comprehensive.