Regulators require that every number printed in a clinical-study report (CSR) be internally consistent across its TFLs and traceable to underlying observations. Scripted Quality Assurance around rich-text outputs (RTF) is brittle under schema drift and provides weak provenance. We present an ontologycentered workflow that converts RTF to JSON, maps JSON to RDF using compact YAML mappers read by a deterministic translator, enforces structure with SHACL, and applies SPARQL rule suites for content checks. Large-language models (LLMs) assist only in drafting YAML; the converter is deterministic and auditable. A pilot across three studies reduced manual TFL QC by up to 75% while surfacing discrepancies earlier.
Relational warehouses can load the tables but they do not encode relationships such as derivesFrom or hasBaselineFlag. They therefore cannot answer cross-artifact questions like “Is the baseline value in the Vital-Sign Table the same as in the Vital-Sign Listing?” Procedural rule engines break when schemas drift and version-controlling thousands of SQL files becomes unmanageable. Column labels alone cannot capture semantics such as “baseline is the last non-missing value before first dose.”
An OWL vocabulary stabilizes concepts (subjects, visits, observations, summary cells), SHACL expresses structural constraints, and SPARQL encodes cross-artifact content rules. Because rules bind to concepts rather than column labels, validations remain stable even when table layouts change.
A layout-aware parser extracts tables/listings from RTF files, normalizes headers, and emits canonical JSON while preserving complete provenance (file id, page, row, column). Figure 1 illustrates the complete ETL workflow.
A deterministic Python translator reads compact YAML mapper files to transform JSON into RDF. The translator mints namespaced IRIs following consistent patterns, performs type casting (e.g., xsd:decimal), threads prov:wasDerivedFrom relationships, and outputs idempotent NTriples suitable for re-runs and difs.
YAML mapper contents: Class instantiation targets, IRI templates, property mappings, and provenance passthrough rules. Only the translator consumes YAML; the triplestore receives the resulting RDF.
YAML creation process: Mappers are drafted from 2–3 representative JSON rows via constrained prompts, then human-reviewed and unit-tested before deployment.
The repository accepts only graphs satisfying predefined shapes. Conformance and violations are recorded as machine-readable validation reports.
The automated suite runs after the SHACL gate and mirrors established QC practice on the graph:
Structure. SHACL ensures well-formedness (typing and cardinalities; mandatory properties present; links such as observes point to Observation).
Content. A library of SPARQL templates re-computes denominators/percentages, verifies cross-table agreement (e.g., listings vs. summary tables), and enforces key business rules (e.g., baseline and change-from-baseline linkages; date constraints for dosing vs. events). Findings are emitted as RDF validation reports with click-through provenance to file/page/row/column.
Where rules come from. Rule templates are derived from sponsor QC checklists/SOPs,
CDISC SDTM and ADaM implementation guides (variables, flags, and derivations), and ICH E3
guidance on CSR structure/consistency [
LLMs (GPT-4, Claude-3) are used only to draft YAML mappers, delivering ∼70% faster authoring; the converter never calls an LLM, keeping ETL deterministic and auditable. One of the many examples of hallucination is Representative hallucination (inconsistent IRIs): • Input: Subject “P001” in multiple rows • LLM: Generates both ex:subject/P001 and ex:subj/P-001 • Impact: Duplicate subjects and incorrect aggregations Why fine-tuning/APIs didn’t fit; why prompt tuning did: public APIs lack fixed seeds, so drafts vary across runs (failing audit reproducibility), while on-prem fine-tuning strained GPUs (quantization degraded output; full precision was impractical). Constrained prompt tuning with a YAML schema and 2–3 golden JSON→triple examples yielded stable drafts; the deterministic translator preserved auditability.
Across three internal studies, manual TFL QC efort decreased by up to 75% with earlier detection of denominator/baseline inconsistencies (internal benchmarks, Q4 2024). Gains stem from removing duplicate programming and enabling click-through lineage for auditors. Regulatory writers gain machine-readable provenance they can paste straight into submission dossiers
Separating semantics (ontology + shapes + rules) from layout (YAML mappers) yields robust, auditable TFL validation for the CSR. The graph layer provides stable checks under schema drift and click-through provenance from any printed number to its contributing rows.
During the preparation of this work, we, the authors, used ChatGPT for grammar and spelling checks, paraphrasing, and citation management. After using this service, we reviewed and edited the content as needed and take full responsibility for the publication’s content.