The Problem
Inclusion of patients in national clinical and tumor registries is essential for quality improvement, patient safety, and risk adjustment at Duke University. However, manual data abstraction is costly and labor-intensive, especially given the current focus on expansion of Duke Health. We propose large language model–driven pipelines to streamline registry data entry, enhance efficiency, and strengthen multi-institutional collaboration.
Our Solution
We will develop LARP: LLM-augmented registry pipelines using DIHI’s Scout to normalize electronic health record data into structured, machine-readable representations, apply prompt-engineered LLM extraction, and convert outputs to registry fields. The human-in-the-loop workflow, modular design, and validation against retrospective registry data ensure accuracy, adaptability, and scalable deployment within Duke Health.
Anticipated Impact
LARP will substantially reduce personnel hours and costs for registry abstraction, potentially halving resource requirements, freeing staff for quality improvement. Accurate, rapid registry population will accelerate multi-institutional research, enable creation of rare disease registries, and strengthen Duke’s leadership in clinical data curation while preserving auditability, human oversight, and regulatory compliance.
