Harnessing Large Language Models in Neonatal Intraventricular Hemorrhage: Exploring Retrieval Augmented Generation...

Area of Science:

• Neonatal Medicine
• Artificial Intelligence
• Medical Informatics

Background:

• Neonatal intraventricular hemorrhage (IVH) is a significant concern in preterm infants, with numerous studies investigating its prognostic variables.
• Accurate prediction of IVH outcomes is essential for timely clinical intervention and improved patient management.
• Existing literature synthesis methods can be time-consuming and may not capture all relevant predictive factors.

Purpose of the Study:

• To assess the capability of large language models (LLMs) to autonomously synthesize literature and extract prognostic variables for neonatal IVH.
• To evaluate LLMs' performance in ranking clinical features and stratifying risk for IVH outcomes.
• To explore the potential of AI in identifying key predictors for mortality, progression, complications, and resolution of IVH.

Main Methods:

• A systematic literature review combined with retrieval augmented generation (RAG) methodology was employed.
• GPT-4 and Claude Sonnet LLMs were utilized to identify and extract data from peer-reviewed studies on IVH prediction in preterm neonates.
• Data extraction followed TRIPOD AI guidelines, with semi-automated RAG extraction and manual validation to ensure accuracy and mitigate hallucinations.

Main Results:

• LLMs identified 28 relevant studies, extracting 14 distinct prognostic predictors across four outcome domains.
• Key predictors identified include gestational age (41%), birth weight (25%), and APGAR scores (34%).
• A preliminary risk stratification model indicated high-risk neonates (<28 weeks, <1000g, APGAR <3) with >70% progression risk and >50% mortality risk.

Conclusions:

• LLMs can effectively synthesize medical literature and extract clinically relevant prognostic variables for neonatal IVH.
• Rigorous clinical oversight and human validation are necessary to address LLM-related hallucinations and ensure data reliability.
• Identified predictors form a basis for AI-assisted clinical decision support tools, with future research needed for complication and resolution prediction.