AN END-TO-END LLM PIPELINE FOR HEALTH ECONOMIC TECHNICAL REPORTS: RIGHTS-VERIFIED SOURCES, EXPERT-APPROVED DRAFTING, AND HALLUCINATION-FREE CITATIONS
Author(s)
Andre Verhoek, MSc1, Suzan Serip, MSc2, Yiduo Zhang, BA, MA, PhD2.
1HMS Lead, AstraZeneca, Barcelona, Spain, 2AstraZeneca, Barcelona, Spain.
1HMS Lead, AstraZeneca, Barcelona, Spain, 2AstraZeneca, Barcelona, Spain.
OBJECTIVES: Three barriers limit large language model (LLM) adoption for health technology assessment (HTA) technical reports: drafting requires 5-15 days of expert effort; LLMs hallucinate bibliographic references; and ingesting copyrighted publications without verified text-and-data-mining (TDM) / GenAI rights creates compliance exposure. We developed and validated an integrated agentic LLM pipeline addressing all three.
METHODS: The pipeline comprises three stages. (1) Rights verification: every source DOI is checked against publisher TDM/GenAI licensing via an automated rights-check service before ingestion; non-compliant sources are flagged for substitution. (2) Report generation: an LLM agent (Claude, Anthropic) extracts inputs, methods, and results from Excel models via static inspection and drafts Word documentation aligned with NICE/ISPOR conventions. (3) Structured citations: a two-file output pairs the .docx report — containing EndNote temporary citations ({Author, Year #N}) — with a companion .ris file holding structured metadata, with all AI-generated fields flagged for human verification. The pipeline was validated across six health economic models (3 CEMs including a 54-sheet partitioned-survival model, 1 BIM, 1 dual-scenario evaluation, 1 statistical analysis report) in two phases with prompt refinement.
RESULTS: Rights: 100% of DOI-bearing sources passed automated verification, with a complete compliance audit trail. Reports: generated in 1.5-8 hours at $2-$7/model versus 5-15 days manual; Phase 1 (n=4) produced 13-17 expert reviewer comments per report; Phase 2 (n=2, refined prompts + Opus 4.6) produced zero expert reviewer comments. Citations: 100% RIS structural validity (importable into EndNote without error), 100% citation-RIS synchronisation, 85-90% metadata accuracy on first generation; reference compilation reduced from 2-4 hours to <10 minutes.
CONCLUSIONS: A single integrated LLM pipeline addresses speed, citation hallucination, and TDM rights compliance — the principal barriers to LLM use in HTA documentation. Separating verifiable structure (RIS, rights log) from generated content yields a deployable, auditable workflow compatible with existing reference managers and compliance processes.
METHODS: The pipeline comprises three stages. (1) Rights verification: every source DOI is checked against publisher TDM/GenAI licensing via an automated rights-check service before ingestion; non-compliant sources are flagged for substitution. (2) Report generation: an LLM agent (Claude, Anthropic) extracts inputs, methods, and results from Excel models via static inspection and drafts Word documentation aligned with NICE/ISPOR conventions. (3) Structured citations: a two-file output pairs the .docx report — containing EndNote temporary citations ({Author, Year #N}) — with a companion .ris file holding structured metadata, with all AI-generated fields flagged for human verification. The pipeline was validated across six health economic models (3 CEMs including a 54-sheet partitioned-survival model, 1 BIM, 1 dual-scenario evaluation, 1 statistical analysis report) in two phases with prompt refinement.
RESULTS: Rights: 100% of DOI-bearing sources passed automated verification, with a complete compliance audit trail. Reports: generated in 1.5-8 hours at $2-$7/model versus 5-15 days manual; Phase 1 (n=4) produced 13-17 expert reviewer comments per report; Phase 2 (n=2, refined prompts + Opus 4.6) produced zero expert reviewer comments. Citations: 100% RIS structural validity (importable into EndNote without error), 100% citation-RIS synchronisation, 85-90% metadata accuracy on first generation; reference compilation reduced from 2-4 hours to <10 minutes.
CONCLUSIONS: A single integrated LLM pipeline addresses speed, citation hallucination, and TDM rights compliance — the principal barriers to LLM use in HTA documentation. Separating verifiable structure (RIS, rights log) from generated content yields a deployable, auditable workflow compatible with existing reference managers and compliance processes.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
EE294
Topic
Economic Evaluation, Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Trial-Based Economic Evaluation
Disease
No Additional Disease & Conditions/Specialized Treatment Areas