Professional · Issue 3 · 2026-10-28
EHR-to-Registry Pipeline: Savana's NLP-Driven Approach to Real-World Evidence Generation
Editorially reviewed research summaries, context, and limitations.
J Marin-Corral, N Iglesias, S Menke, et al.
Summary: This article describes the Savana Next-Generation Registry (SNGR), a methodology for constructing oncology-focused real-world data registries by extracting and structuring clinically relevant information from unstructured electronic health record (EHR) text using a multilingual natural language processing (NLP) engine called EHRead®. The system incorporates medical expert curation, terminology standardization, and a quality assurance framework intended to support reproducibility and regulatory compliance. To date, the platform has supported more than 65 real-world use cases and contributed to over 30 peer-reviewed publications across multiple therapeutic areas.
Why it matters: A substantial proportion of clinically meaningful information in EHRs resides in free-text narratives—clinical notes, discharge summaries, pathology reports—that remain inaccessible to conventional structured-data queries. Scalable NLP-based extraction that converts this information into analysis-ready datasets could meaningfully expand the scope and depth of real-world evidence, particularly in settings where staging, treatment response, and disease progression are often documented narratively. Alignment with European regulatory frameworks including the European Health Data Space and GDPR is operationally relevant as regulatory bodies increasingly consider RWE in decision-making.
Expert perspective: This paper is best characterized as a methods and infrastructure description rather than a primary research study, and it should be read accordingly. The abstract provides no head-to-head validation data—no systematic comparison of NLP-extracted variables against manual chart review or a reference standard across a defined sample. The claim of "methodologically robust infrastructure" therefore rests largely on the authors' own characterization and the volume of downstream use cases, neither of which constitutes independent validation. Thirty peer-reviewed publications are cited as evidence of the platform's utility, but publication count is a weak surrogate for data quality; it reflects adoption more than accuracy.
Key questions go unanswered in the abstract: What are the precision, recall, and F1 performance metrics of EHRead® across clinically critical variables? How does extraction performance vary across languages, institutions, and EHR systems? What proportion of records have sufficient free-text richness to yield complete, analysis-ready datasets, and how is missingness handled? The multilayer quality assurance framework is mentioned but not described in evaluable detail.
That said, the underlying problem the platform addresses is genuine and important. Structured EHR fields alone are demonstrably insufficient for complex RWE generation, and NLP-based enrichment is a legitimate and increasingly mature approach. The regulatory alignment work—particularly with the European Health Data Space—represents real operational value for multicenter and cross-border studies. For small patient populations where extracting maximum information from existing clinical documentation is essential, analogous infrastructure could prove highly relevant. Readers should treat this as a credible description of a functional platform with meaningful real-world deployment, while recognizing that independent, prospective validation of extraction accuracy across diverse clinical settings remains the critical unmet evidentiary standard.
Relevance: Registry and Real-World Data Methods
Source: https://pubmed.ncbi.nlm.nih.gov/42830900/
---