Large language models for structured cardiovascular data extraction: a foundation for scalable research and clinical...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Automated information extraction from cardiac reports is crucial for clinical reporting and research.
• Large language models (LLMs) show potential for automating this process, but their clinical performance and implementation across different computing environments require evaluation.

Purpose of the Study:

• To assess the feasibility and performance of LLM-based classification of echocardiogram and invasive coronary angiography reports.
• To evaluate LLM performance on real-world clinical data across local, high-performance computing, and cloud-based platforms.

Main Methods:

• Cardiac reports (n=1000) for acute coronary syndrome patients were labeled for key diagnostic elements (LVF, culprit vessel, acute occlusions).
• LLM-based classification models were developed using prompt-based and fine-tuning approaches.
• Performance was evaluated across model types and compute infrastructures, considering class imbalance and implementation costs.

Main Results:

• LLMs demonstrated strong performance in extracting structured diagnostic information from cardiac reports.
• Cloud-based models (e.g., GPT-4o) achieved high accuracy (0.87 for culprit vessel, 1.0 for LVF).
• Local high-performance cluster models also showed reasonable accuracy (0.634 for culprit vessel, 0.984 for LVF), especially for simpler tasks.

Conclusions:

• LLMs can reliably classify structured cardiology reports across diverse computing infrastructures.
• Their accuracy and adaptability support integration into clinical workflows and research pipelines for scalable report structuring and dataset generation.