Related Experiment Video
Updated: Sep 14, 2025
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
684
Automated transformation of unstructured cardiovascular diagnostic reports into structured datasets using
Sumukh Vasisht Shankar1, Lovedeep S Dhingra1, Arya Aminorroaya1
1Section of Cardiovascular Medicine, Department of Internal Medicine, Yale School of Medicine, 195 Church Street, 6th Floor, New Haven, CT 06510, USA.
European Heart Journal. Digital Health
|July 24, 2025
Summary
Large language models (LLMs) can now extract valuable cardiovascular data from unstructured echocardiogram reports, improving data accessibility for clinical insights and research.
Area of Science:
- Cardiovascular diagnostics
- Artificial intelligence in medicine
- Natural Language Processing
Background:
- Cardiovascular diagnostic testing generates rich data often trapped in unstructured reports.
- Limited accessibility of this data hinders its clinical utility and research potential.
Purpose of the Study:
- To develop and validate a novel approach using large language models (LLMs) to transform free-text echocardiographic reports into structured, tabular datasets.
- To improve the accessibility and usability of cardiovascular data.
Main Methods:
- Sequentially deployed generative and interpretative open-source LLMs (Llama2-70b, Llama2-13b).
- Fine-tuned Llama2-13b (HeartDX-LM) using generated transthoracic echocardiogram (TTE) reports for data extraction across 18 echocardiographic fields.
- Evaluated HeartDX-LM on diverse datasets, including Yale New Haven Health System (YNHHS), MIMIC-III, and MIMIC-IV.
Main Results:
- HeartDX-LM achieved high accuracy in data extraction: 98.7% at YNHHS, 87.1% on older YNHHS reports, 87.9% in MIMIC-III, and 91.3% in MIMIC-IV.
- Identified a minimum of 500 unstructured reports-structured data pairs for effective fine-tuning.
- The HeartDX-LM model has been publicly released.
Conclusions:
- A novel approach using paired LLMs effectively transforms free-text echocardiographic reports into valuable tabular datasets.
- This method significantly enhances the accessibility and utility of cardiovascular data for clinical applications and research.
