Retrieval-Augmented Large Language Model for Angiographic Prediction of Coronary Physiology.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Cardiovascular Physiology
• Machine Learning for Healthcare

Background:

• Invasive physiologic assessment is crucial for moderately stenotic coronary lesions but underutilized.
• The capability of large language models (LLMs) to estimate coronary physiology from angiographic images is unexplored.
• The potential of retrieval augmentation to enhance LLM performance in this task is unknown.

Purpose of the Study:

• To evaluate if GPT-based LLMs can estimate coronary physiology (instantaneous wave-free ratio [iFR]) from coronary angiographic images.
• To determine if retrieval augmentation improves the accuracy of LLM-based iFR estimation.
• To assess the agreement between predicted and invasively measured iFR.

Main Methods:

• Retrospective pilot study involving 32 vessels with invasive iFR assessment and two orthogonal angiographic frames.
• Comparison of a baseline GPT-5.2 model (No-RAG) with a GPT-5.2 model enhanced with retrieval-augmented generation (RAG).
• RAG incorporated case-specific text chunks from relevant documents to provide context without additional lesion-specific data.

Main Results:

• The RAG-enhanced model significantly improved agreement between predicted and measured iFR.
• Mean absolute error decreased from 0.064 (No-RAG) to 0.029 (RAG); root mean square error decreased from 0.083 to 0.038.
• Correlation between predicted and invasive iFR increased substantially from r = 0.205 (No-RAG) to r = 0.830 (RAG).

Conclusions:

• Retrieval augmentation enhanced GPT-based LLM performance in estimating iFR from coronary angiographic images.
• The findings suggest potential for AI in functionally classifying coronary stenoses, though overestimation in less severe lesions requires a scaling correction.
• Validation in larger, more balanced cohorts is necessary to confirm these promising results.