Transforming free-text coronary angiography reports into structured, analyzable data using large language models

Abstract

Coronary angiography (CAG) reports contain many details about coronary anatomy, lesion characteristics, and interventional procedures. However, their free-text format limits their research utility. Therefore, we sought to develop and validate a framework leveraging large language models (LLMs) to convert CAG reports automatically into a standardized structured format. Using 50 CAG reports from a tertiary hospital, we developed a multi-step framework to standardize and extract key information from CAG reports. First, a standard annotation schema was developed by cardiologists. Thereafter, an LLM (GPT-4o) converted the free-text CAG reports into the hierarchical annotation schema in a standardized format. Finally, clinically relevant information was extracted from the standardized schema. One hundred CAG reports from each of two hospitals were used for internal and external test, respectively. The 12 key information points included four CAG-related (previous stent information, lesion characteristics, and anatomical diagnosis) and eight percutaneous coronary intervention (PCI)-related key points (complex PCI criteria and current stent information). For internal test, two interventional cardiologists independently extracted information, with discrepancies resolved through consensus, as reference standard. Based on the reference standard, the proposed framework demonstrated superior accuracy for CAG-related (99.5% vs. 91.8%; p < 0.001) and comparable accuracy for PCI-related key points (98.3% vs. 97.4%; p = 0.512) in the internal test. External test confirmed high accuracy for both CAG- (96.2%) and PCI-related key points (99.4%). This framework demonstrated excellent accuracy in standardizing free-text CAG reports, potentially enabling more efficient utilization of detailed clinical data for cardiovascular research.