CBR-to-SQL rethinks Text-to-SQL for EHRs with case-based reasoning

What’s new

A new arXiv preprint, “CBR-to-SQL: Rethinking Retrieval-based Text-to-SQL using Case-based Reasoning in the Healthcare Domain,” proposes a twist on retrieval-augmented generation for medical databases: use case-based reasoning to guide Large Language Models (LLMs) from clinical questions to precise SQL. Instead of retrieving only schema notes or generic documentation, the method stores and retrieves prior “cases”—pairs of natural-language queries and validated SQL—then adapts the closest matches to fit the current schema and intent. The goal is straightforward: reduce hallucinations, respect complex medical ontologies, and make LLM outputs executable on real Electronic Health Record (EHR) systems.

Why it matters

EHR data is locked behind intricate schemas, idiosyncratic table names, and strict privacy constraints. Clinicians and analysts often lack SQL fluency, while end-to-end LLMs can drift on domain-specific joins, time windows, or cohort definitions. Retrieval-based text-to-SQL has emerged as a practical bridge, but healthcare demands higher precision and auditability. Case-based reasoning—an old-school AI idea revived for the LLM era—offers a traceable path: “find a similar case, then adapt it.” In a regulated setting, that transparency can be the difference between a safe decision support tool and an untrusted black box.

China and the world

The approach is well-timed for health systems navigating data localization and privacy regimes such as HIPAA in the US and China’s Personal Information Protection Law (PIPL). Hospitals in China commonly run heterogeneous Hospital Information Systems, where schemas vary by vendor and region; a case-driven adapter could temper that fragmentation without mass retraining. It also dovetails with on-prem deployments favored amid US export controls on advanced AI chips, which push providers toward lighter models paired with smarter retrieval. Chinese AI leaders—Baidu (百度), Alibaba (阿里巴巴), Tencent (腾讯), and iFlytek (科大讯飞)—have all touted medical LLMs; methods that lean on local, auditable case libraries could make those systems more reliable in production.

What to watch

Key questions remain. How robust is case selection when queries are novel or when hospital schemas evolve? Can institutions safely curate and share de-identified case libraries across sites to reduce cold-start issues, or will data governance keep them siloed? And does case-guided adaptation consistently outperform naive RAG or fine-tuned text-to-SQL baselines across medications, labs, imaging, and longitudinal cohorts? As the preprint lands on arXiv, the next step is rigorous, real-world evaluation—where execution accuracy, privacy posture, and maintenance cost will ultimately decide whether CBR-to-SQL makes it from paper to ward.