A Feasibility Study of Literature-Guided HRV Stratification Using Large Language Models

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine

Background:

• Heart rate variability (HRV) is crucial for vascular health assessment.
• Clinical decision support systems (CDSSs) struggle to keep pace with evolving HRV literature.
• Systematic literature synthesis is needed for accurate HRV-based risk stratification.

Purpose of the Study:

• To develop an LLM-assisted framework for synthesizing HRV literature.
• To support transparent risk stratification using HRV evidence.
• To enable systematic extraction and organization of HRV data from studies.

Main Methods:

• An LLM-driven framework extracted HRV parameters from 140 medical abstracts.
• The system simulated human reasoning for identifying HRV indicators and grouping patient data.
• Performance was evaluated using ECG-derived HRV features for literature-guided classification.

Main Results:

• The framework achieved 86% accuracy, 81% sensitivity, and 87% specificity in HRV classification.
• The LLM-assisted system offered transparent, literature-grounded reasoning.
• It demonstrated adaptability to new research, unlike traditional machine learning.

Conclusions:

• LLMs can support evidence-guided parameter selection for HRV risk stratification.
• This approach enhances transparency and addresses 'black box' concerns in AI-assisted CDSS.
• LLMs reduce manual effort in clinical decision support development.