Presentation (CTI)

OBJECTIVES: Budget-impact models (BIMs) submitted to health technology assessment (HTA) agencies are typically implemented in Excel, where undetected structural errors can affect reimbursement decisions. Although large language models (LLMs) can interpret model content, they often lack awareness of the model-building process required for quality control (QC). Such awareness requires structural understanding of the model, including the expected placement of model elements, the way formulas operationalize epidemiologic concepts, and how interconnected components, hidden controls, and assumptions generate outcomes. We developed XLGEN, a structure-first framework that converts BIMs into knowledge graphs to support auditable and QC-oriented AI reasoning.
METHODS: XLGEN converts Excel-based BIMs into knowledge graphs through a two-stage process. First, structural ingestion identifies tables, extracts formula dependencies, and reconstructs workbook organization as a directed graph, where cells are nodes and formula references are edges. Second, a domain layer grounds graph objects against BIM ontology rules covering comparators, epidemiology, and cost outcomes. The resulting graph is queried through an agentic retrieval layer that traverses dependencies, validates ontology constraints, and requires every response to cite a specific cell (Sheet!Cell). The framework was evaluated on three HTA-grade workbooks using five predefined QC tasks, including comparator identification, population tracing, input-parameter localization, and cost-driver analysis. Citation-grounded accuracy was compared against Claude Opus 4.7 and GPT-5.5.
RESULTS: Across three workbooks, XLGEN answered 95% of benchmark questions with valid Sheet!Cell citations, versus 89% for Claude Opus 4.7 and 85% for GPT-5.5. XLGEN also flagged 12 structural inconsistencies, including broken dependencies and ontology-rule violations, that both chatbots missed. Every accepted answer remained traceable to the underlying cells, enabling reviewer verification.
CONCLUSIONS: Representing an Excel BIM as a knowledge graph preserves structural and domain knowledge embedded in model design, which in turn, enables AI reasoning to be aligned with HTA verification practices. Structure-first, ontology-grounded frameworks are essential for trustworthy AI-assisted QC of BIMs.