316 - Harnessing large language models for literature synthesis and prognostic variable discovery in intraventricular hemorrhage

Publication Number: 4310.316

Tanima Arora, Boston Children's Hospital, Boston, MA, United States; Kristyn S. Beam, Beth Israel Deaconess Medical Center, Needham, MA, United States

Background: Intraventricular hemorrhage (IVH) is a major complication of prematurity, particularly in very low birth weight infants, with mortality ranging from 4-40% depending on severity. Prognostication currently relies on clinical gestalt and basic demographic factors. Large language models (LLMs), such as generative AI tools, offer novel opportunities to enhance prediction of neonatal IVH outcomes.

Objective: To apply LLMs for literature synthesis and prognostic variable discovery in IVH, focusing on mortality, disease progression, complications (post-hemorrhagic hydrocephalus), and resolution at 30 days or discharge.

Design/Methods: Claude 4.0 Sonnet and GPT-5.0 were used to conduct a systematic review of studies employing machine/deep learning for IVH outcomes, following TRIPOD-AI guidelines. Full texts were ingested into a retrieval-augmented generation (RAG) framework, which embedded documents into vector space and enabled targeted querying of outcome-specific predictors. Iterative prompt engineering guided the models to extract structured data (e.g., predictor name, frequency, outcome association) and resolve synonymous variables into standardized terms. Extracted outputs were then cross verified across both LLMs to improve consistency and reduce hallucination.

Results: Thirty-one relevant studies were identified: mortality (9; 29%), progression (16; 52%), complications (5; 16%), and none on IVH resolution, underscoring a critical research gap. Variable extraction revealed consistently high-impact predictors: gestational age (35%), birth weight (19%), and Apgar scores (23%). Outcome-specific predictors included respiratory distress syndrome and IVH grade for mortality, platelet count and birth location for progression, and mean arterial pressure for complications. Risk profiling showed that infants with GA < 28 weeks, BW < 1000 g, and low Apgar scores had poor outcomes across categories, while moderate/low-risk groups displayed outcome-specific variables, suggesting utility for AI model training.

Conclusion(s): This is the first systematic application of LLMs to neonatal IVH prognostication. The absence of studies on resolution prediction highlights an unmet clinical need, while the identified variable hierarchy provides a foundation for future AI-driven prognostic models. Such approaches could enable earlier intervention and more precise care planning for high-risk neonates.

Impact Stratification of Key Prognostic Variables
Figure 1. Impact stratification of key prognostic variables for neonatal IVH outcomes (mortality, progression and complications) extracted via a RAG pipeline by large language models (Claude 4.0 & GPT 5.0). A) Universal high impact variables consistent amongst all neonatal IVH outcomes. B) Outcome specific high impact prognostic variables. C) Modifiable risk prognostic variables D) Composite clinical impact ranking of LLM-extracted prognostic variables.

LLM-Extracted Key Predictors for Neonatal Outcomes
Figure 1. Top predictive variables for neonatal IVH outcomes, including a mortality b) progression c) development of complications and d )universal predictors across all outcomes, as extracted by the large language models using a RAG pipeline.