Automated Transformation of Unstructured Cardiovascular Diagnostic Reports into Structured Datasets Using...

Area of Science:

• Cardiovascular medicine
• Artificial intelligence
• Medical informatics

Background:

• Cardiovascular diagnostic testing generates rich data often trapped in unstructured reports.
• Manual data abstraction limits real-time patient care and research applications.
• Automating data extraction from echocardiogram reports is crucial for clinical utility.

Purpose of the Study:

• To develop and validate a novel method for automated data extraction from unstructured transthoracic echocardiogram (TTE) reports.
• To leverage large language models (LLMs) for transforming free-text clinical narratives into structured, computable data.
• To improve the accessibility of cardiovascular data for patient care and research.

Main Methods:

• A two-step process using generative (Llama2 70b) and interpretative (Llama2 13b) LLMs was employed.
• Llama2 70b generated varied TTE report formats from real-world data.
• Llama2 13b was fine-tuned to extract 18 key echocardiographic fields from free-text narratives, creating the HeartDx-LM model.

Main Results:

• The HeartDx-LM model achieved high accuracy in extracting data from contemporary (98.7%) and older (87.1%) echocardiogram reports.
• External validation on MIMIC-III (87.9%) and MIMIC-IV (91.3%) datasets demonstrated robust performance.
• Stable performance was achieved with as few as 500 annotated reports for fine-tuning.

Conclusions:

• A novel method using paired LLMs automates the extraction of unstructured echocardiographic reports into tabular datasets.
• This scalable strategy transforms unstructured reports into computable elements.
• The approach promises to enhance cardiovascular care quality and facilitate research.