CAVE-Onc: Graph-constrained agentic validation for cross-domain contradictions in CDISC oncology submissions

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). CAVE-Onc operates over an RDF knowledge graph built from XPT datasets. Layer 1 comprises 111 SHACL shapes—85 ported from CORE rules, 8 RECIST 1.1 derivation shapes, and 18 archetype-specific SHACL-SPARQL cross-domain constraint shapes—while Layer 3 implements a CaveAgent for RECIST Table 7 verification. We evaluated CAVE-Onc against pre-registered hypotheses (OSF): Track A assessed complementarity with CORE on clean data (52 subjects); Track B evaluated detection of 20 contradiction archetypes injected via a RELREC-preserving generator. On Track A, CAVE L1 and CORE produced disjoint flag sets (Jaccard = 0.004). On Track B, CAVE detected 20/20 archetypes (100%; 95% CI: 83.2%–100.0%): 19 via L1 SHACL-SPARQL constraints and 1 via L3 agent exclusively. Two industry validators—the CDISC CORE engine (8/20) and the Pinnacle 21 FDA engine (6/20)—each detected 0/10 of the cross-domain RECIST contradictions they cannot express (all detections were CORE-seeded structural checks), versus CAVE’s 10/10 (McNemar p = 0.002). Three domain experts validated all 20 archetypes (Fleiss’ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" id="equation1"> <mml:mrow> <mml:mi>κ</mml:mi> <mml:mo>=</mml:mo> <mml:mn>0.705</mml:mn> </mml:mrow> </mml:math> ; 0 rated invalid). Because the archetypes were authored knowing the injected patterns, we frame this as construction validation of expressiveness , not a real-world detection estimate; on five held-out archetypes, existing shapes detected 3/5. On two real oncology trials mapped to SDTM, the engine stayed specific on unmutated data and detected 10/11 and 16/18 of applicable archetypes. Graph-based validation thus augments rather than replaces industry-standard tools.