Emergency medical service activity logs represent vast repositories of public-sector operational records, but combining structured dispatch databases with narrative evaluation notes has persistently challenged analytical workflows. To address this operational gap, researchers Hyejin Kim, Chaehee Park, and Youngjung Suh developed and evaluated a system-level framework combining natural language–to–SQL query generation and retrieval-augmented generation. Rather than presenting isolated algorithmic improvements, the architecture unifies structured statistical querying, automated structuring of clinical narratives, metadata-aware case retrieval, and policy synthesis within a single workflow.
Operational records in emergency medical systems naturally bifurcate into two distinct data types: tabular dispatch information detailing call parameters and unstructured prose written during patient assessments. The research team configured their framework to link these formats systematically, enabling users to interrogate statistical metrics while contextualizing operational trends with relevant historical cases. The system-level pipeline incorporates four foundational stages designed to maintain continuous analytical context across successive interactions.
In the first functional stage, the system converts natural-language questions into executable database queries via structured statistical query generation. To capture operational nuances, the evaluation scheme accounts for runtime execution and structural query diagnostics alongside generation failures and empty-result returns. The second phase applies generative artificial intelligence to parse unstructured evaluation notes into a standardized 5W1H narrative representation. This structuring organizes who, what, when, where, why, and how elements from free-text evaluations before retrieval occurs.
The third component performs case-based search across unstructured text using retrieval-augmented generation coupled with metadata filtering. In testing chunking strategies and hybrid retrieval configurations, the researchers measured retrieval quality using Filter Satisfaction Rate, Filter Drop Rate, and RAGAS-based metrics evaluating faithfulness and relevance. The experiments demonstrated noticeable trade-offs across dense, sparse, and metadata-aware retrieval pipelines, with metadata filtering delivering its most prominent retrieval performance advantages on explicitly metadata-relevant search tasks.
The concluding stage integrates accumulated evidence into a synthesis module designed for evidence-based policy formulation. In policy control evaluations, providing accumulated analytical context consistently increased Evidence Support and Evidence Coverage compared to closed-book generative models. When compared against single-turn hybrid retrieval setups, accumulated analytical evidence reliably elevated Evidence Coverage, although it did not demonstrate a corresponding increase in Evidence Support across evaluated scenarios.
Controlled experimental evaluations across 36 structured test queries highlighted important constraints in query formulation. Specifically, the data provided no statistically supported improvements in query execution success from runtime-introspected schema representations or fixed three-shot prompting strategies. Furthermore, an independent expert evaluation revealed criterion-dependent reliability across outputs and showed an uncertain association between Evidence Quality and Overall Quality ratings.
Overall, the empirical findings demonstrate the technical feasibility of bridging structured public-sector emergency metrics and unstructured narrative logs within evaluated scenarios. While the framework establishes an empirical foundation for multi-stage decision support, the observed component trade-offs emphasize the necessity of rigorous failure diagnostics and targeted retrieval filtering when utilizing combined analytical models.
Subject of Research: Integrated NL2SQL and RAG framework for EMS log analytics
Article Title: An integrated NL2SQL–RAG framework for large-scale EMS log analytics and evidence-based policy analysis
Article References: Kim, H., Park, C., & Suh, Y. (2026). An integrated NL2SQL–RAG framework for large-scale EMS log analytics and evidence-based policy analysis. Journal of Big Data. https://doi.org/10.1186/s40537-026-01621-3
Image Credits: AI Generated
DOI: 10.1186/s40537-026-01621-3
Keywords: emergency medical services, NL2SQL, retrieval-augmented generation, data analytics, public sector data, dispatch records, evaluation notes, 5W1H representation, hybrid retrieval, metadata filtering, Filter Satisfaction Rate, Filter Drop Rate, RAGAS metrics, evidence-based policy, query execution, decision support
News Source: Denise Maddox. (November 20, 2026). Integrated Framework Connects Structured Analytics and Case Retrieval for EMS Logs. Scienmag.



