Related Experiment Video
Updated: Jun 13, 2026
Upper-extremity Approach for Secondary Access in Transfemoral Transcatheter Aortic Valve Implantation
Published on: August 8, 2025
From Theory to Practice: GPT-Supported Data Extraction in Observational Studies on Transcatheter Aortic Valve
Gloria Brigiari1, Roberta Dotto2, Carlo Cernetti2
1Unit of Biostatistics, Epidemiology and Public Health, Department of Cardiac, Thoracic, Vascular Sciences and Public Health, University of Padova, Padova, Italy; BIOSTAT-X Biostatistics & AI for Biomedical Discovery, Pediatric Research Institute (IRP) "Città della Speranza", Padova, Italy; PhD Program in Translation Specialistic Medicine "G.B. Morgagni", Curriculum "Biostatistics and Clinical Epidemiology", University of Padova, Padova, Italy.
Generative pre-trained transformer (GPT) models show high accuracy in extracting data from electronic health records for transcatheter aortic valve replacement studies. This demonstrates the feasibility of using large language models (LLMs) to streamline observational research.
Area of Science:
- Cardiology
- Medical Informatics
- Artificial Intelligence in Healthcare
Background:
- Manual data abstraction from electronic health records (EHRs) is laborious, error-prone, and difficult to scale.
- Large language models (LLMs) offer a potential solution to automate data extraction from unstructured EHRs.
- Limited real-world clinical research exists on LLM performance for EHR data abstraction.
Purpose of the Study:
- To evaluate the performance of a generative pre-trained transformer (GPT)-based LLM.
- To assess GPT's accuracy in extracting sociodemographic, procedural, and outcome variables from free-text EHRs.
- Focus on patients undergoing transcatheter aortic valve replacement (TAVR).
Main Methods:
- Retrospective analysis of 108 TAVR cases (January-June 2024) at Ca' Foncello Hospital.
- Manual abstraction by two reviewers served as the reference standard.
- Calculated accuracy, sensitivity, specificity, and used Bland-Altman analysis for continuous variables.
Main Results:
- GPT achieved high accuracy (0.657-1.00) across various variables, including gender, procedure timings, and intraoperative complications.
- Sensitivity reached 1.00 for rare events like intraoperative neurological complications.
- Specificity exceeded 0.90 for most variables; Bland-Altman analysis showed minimal bias for vital parameters.
Conclusions:
- GPT-based data extraction demonstrates high accuracy for continuous measurements and rare intraoperative outcomes in TAVR.
- Performance was lower for infrequent postoperative events due to sparse true positives.
- LLM-assisted extraction is feasible for observational research, warranting further validation in larger cohorts.
