From unstructured oncology EHRs to regulatory-grade real-world evidence: A next-generation registry approach.

Abstract

e23425 Background: Real-world evidence (RWE) derived from oncology electronic health records (EHRs) is increasingly used to inform clinical research, healthcare decision-making, and regulatory discussions. However, most clinically relevant oncology information remains embedded in unstructured free text, limiting scalability, data quality, and reuse. Methods: We implemented the Savana Next-Generation Registry (SNGR), a multilingual software-as-a-service platform designed to generate clinical information from routine EHR data. Participating sites undergo standardized feasibility assessment, source-level pseudonymization, and predefined data quality checks. Heterogeneous EHR data are harmonized into a common data model and curated through automated multidimensional quality controls. Free-text narratives are structured using EHRead, a clinical natural language processing (cNLP) engine trained exclusively on real-world EHRs and integrating international medical terminologies enriched with real-world clinical variants. Terminology panels define key Population–Intervention–Outcome variables. A multilayer clinical quality assurance framework combining automated evaluation and human-in-the-loop validation ensures accuracy, reproducibility, and regulatory reliability. Results: SNGR has enabled over 65 real-world studies across multiple therapeutic areas, including oncology, generating more than 30 peer-reviewed publications. In oncology, SNGR has supported large multicenter registries involving millions of patients, enabling descriptive, comparative, and predictive analyses using routine clinical data. The platform has processed more than 550 million EHRs from over 33 million patients, supporting longitudinal, high-throughput data extraction. Embedded validation workflows consistently reported high precision, recall, and F1-scores for clinically relevant variables. Applications include characterization of treatment patterns and outcomes, prediction of thromboembolic and bleeding risk, and development of explainable prognostic models outperforming traditional staging systems. Conclusions: Next-generation registries based on validated cNLP enable scalable, reliable reuse of unstructured oncology EHR data to generate high-quality, regulatory-aligned RWE.