CAVE-Onc: Graph-constrained agentic validation for cross-domain contradictions in CDISC oncology submissions
Jaime Yan
Abstract
The CDISC Rules Engine (CORE) provides industry-standard validation for Study Data Tabulation Model (SDTM) submissions, yet its imperative, domain-scoped architecture cannot express cross-domain contradictions—such as a Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 overall response that contradicts the combined target lesion, non-target lesion, and new-lesion status—which pass structural validators undetected. We present CAVE-Onc, a two-layer graph-constrained agentic validation engine that augments CORE with declarative Shapes Constraint Language (SHACL) validation (L1) and a LangGraph-based agentic orchestrator (L3).
Introduction
Clinical trial data submitted to regulatory authorities must conform to the Study Data Tabulation Model (SDTM) standards maintained by the Clinical Data Interchange Standards Consortium (CDISC) [1]. The CDISC Rules Engine (CORE) [2] provides an open-source, imperative validation framework that evaluates domain-scoped rules (field lengths, controlled terminology, required variables) encoded in YAML/JSON definitions sourced from the CDISC Library.
Materials and method
Ethics statement
This study analyzed only publicly available, de-identified and synthetic datasets—the CDISC SDTM/ADaM Pilot Project data and the open-source pharmaversesdtm package [7]—and did not involve human participants, identifiable patient data, or the collection of new data. Accordingly, institutional review board (IRB) approval and informed consent were not applicable to this work.
Results
On clean reference data (52 subjects from CDISC Pilot 1, oncology domains DM and EX), CAVE L1 and CORE produced substantially different flag sets (Table 2).
Discussion
Our results demonstrate two complementary findings. First, CAVE-Onc and CORE operate on disjoint rule classes (Jaccard = 0.004), confirming that graph-based validation augments rather than replaces industry-standard tools. Second, CAVE-Onc achieves 100% detection of all 20 clinician-reviewed contradiction archetypes—19 via SHACL-SPARQL constraints and 1 via the L3 agentic orchestrator—demonstrating that the graph-constrained architecture can express the full range of cross-domain clinical contradictions tested.
Conclusion
CAVE-Onc shows that graph-constrained validation can close a clinically important expressiveness gap in oncology SDTM review. Across a pre-registered two-track evaluation, CAVE-Onc remained complementary to CORE on clean reference data and detected all 20 clinician-reviewed contradiction archetypes in the injected corpus—a construction validation of the architecture’s expressiveness rather than an estimate of real-world detection—including one RECIST Table 7 overall-response contradiction handled exclusively by the L3 agent.
Acknowledgments
Three domain experts provided independent archetype validity ratings and explainability assessments used in the G8/G9 evaluation.
Citation: Yan J (2026) CAVE-Onc: Graph-constrained agentic validation for cross-domain contradictions in CDISC oncology submissions. PLoS One 21(8): e0350376. https://doi.org/10.1371/journal.pone.0350376
Editor: Le Zhang, Sichuan University, CHINA
Received: May 11, 2026; Accepted: July 30, 2026; Published: August 14, 2026
Copyright: © 2026 Jaime Yan. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All code, SHACL shapes, evaluation scripts, and synthetic benchmark data are publicly available without restriction in the CAVE-Onc-Benchmark repository (GitHub: yanmingyu92/CAVE-Onc-Benchmark), https://github.com/yanmingyu92/CAVE-Onc-Benchmark. All datasets used in this study are open-source or synthetic (pharmaversesdtm, CDISC SDTM/ADaM Pilot Project). The de-identified per-archetype expert-rating matrix underlying the inter-rater agreement statistic is included in the repository (eval/expert_ratings_deidentified.csv), together with a reproduction script (scripts/reproduce_kappa.py); no reviewer-identifying information is shared. The OSF pre-registration protocol is available in the OSF project osf.io/fx2ky, https://osf.io/fx2ky.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
