Presentation (CTI)

OBJECTIVES: Real-world evidence (RWE) generation is often constrained by the limited clinical granularity of administrative databases and disease registries. Hospital information systems (HIS) hold rich patient-level clinical data, but it is fragmented across heterogeneous systems and largely stored as unstructured text. We developed a scalable methodological framework for generating research-ready clinical databases from routinely collected hospital data by combining structured data extraction and natural language processing (NLP).
METHODS: The framework supports RWE studies for the pharmaceutical and medical-device industries, scientific societies, and healthcare institutions. It comprises five steps: (1) defining research objectives and clinically relevant variables; (2) assessing data availability and interoperability across HIS; (3) developing a standardized data dictionary; (4) classifying each variable as directly extractable from structured sources or requiring AI-assisted NLP from free-text records; and (5) evaluating the technical and organizational feasibility of semi-automated collection. Particular attention is paid to ensuring reproducibility, traceability, and data quality throughout the extraction process.
RESULTS: Across multiple therapeutic areas, standardized data dictionaries were developed covering more than 20 clinical entities and over 150 primary variables — spanning clinical characteristics, care pathway, treatment sequences, etc. Each variable was systematically mapped to either a structured source or AI-assisted NLP extraction from unstructured documentation. NLP performance is evaluated against expert manual annotation using precision, recall, and F1-score. Compared with conventional electronic case report form (eCRF)-based chart abstraction, the framework is designed to reduce manual workload, improve data completeness and consistency, lower inter-observer variability, and enable scalable, reproducible collection of high-quality real-world data.
CONCLUSIONS: This AI-assisted framework addresses a key methodological barrier to RWE generation by turning routinely collected hospital data into research-ready clinical databases. By coupling standardized data models with NLP-based extraction of free-text information, it can improve the efficiency, quality, reproducibility, and scalability of observational studies and health technology assessment across therapeutic areas.