Accelerating clinical research through LLMs: data extraction from nursing diaries in TAVI care

Abstract Background Retrospective clinical research often involves time-consuming structuring of nursing and medical notes. In procedures like Transcatheter Aortic Valve Implantation (TAVI), where longitudinal care is complex and recorded in free text, routinely collected data are vital to understanding care pathways and outcomes. This study explores the potential of Large Language Models (LLMs) to accelerate research and reduce the manual burden in evidence generation. Methods This is a retrospective study of 108 anonymized nursing/medical diaries from TAVI patients treated between January and June 2024 in a cardiology unit in the Veneto Region (Italy). A clinician manually classified free-text data across multiple domains (e.g., vitals, procedural details, complications), as reference. Using GPT-4o (via the OpenAI R package), we prompted the model to extract comparable structured variables. Accuracy, sensitivity, and specificity were key outcomes, also assessing model's ability to handle repeated measures and infer derived metrics (e.g., length of stay). Results LLM accuracy was excellent across most variables (e.g., 100% for gender, Barthel index at admission). Vital signs (e.g., HR, blood pressure) achieved >98% accuracy despite variable frequency and formatting. Derived values (e.g., procedure duration) were also correctly computed. Lower performance was observed for some complications (e.g., postoperative arrhythmias, sensitivity 0.72), reflecting complexity and ambiguity in documentation. LLMs substantially reduced data processing time compared to manual abstraction. Conclusions LLMs can accelerate retrospective research by transforming free-text records into ready-to-analyze datasets. Their ability to handle multilingual data and derive implicit variables from fragmented text supports scalability across healthcare settings. LLMs can reduce barriers for nursing-led research and foster data-driven improvement of clinical pathways, also in complex interventions like TAVI. Key messages • LLMs can turn routine clinical texts into structured research-ready data, cutting manual processing time. • This AI-driven approach enables timely, scalable retrospective research in complex care settings.