Presentation (CTI)

OBJECTIVES: Claims databases lack the clinical granularity (tumor staging, liver function, radiologic response) needed for oncology health technology assessment and reassessment (HTA/HTAR). We evaluated whether hospital electronic medical record (EMR) data, refined using large language models (LLMs), can generate decision-grade real-world evidence, using first-line systemic therapy for hepatocellular carcinoma (HCC) as a case study.
METHODS: Using the multi-institutional Chang Gung Research Database, we applied a validated LLM pipeline to extract high-granularity variables (e.g., tumor burden, tumor size, and tumor number) from unstructured imaging text reports. These variables were combined with structured EMR data within a target trial emulation framework that compared first-line atezolizumab plus bevacizumab (atezo-beva) versus sorafenib in a simulated HTA process (atezo-beva establishment). A comparison of atezo-beva versus lenvatinib was used in the HTAR process (atezo-beva reimbursement). We emulated an IMbrave150-like trial that included Child-Pugh A patients with macrovascular invasion or extrahepatic spread. Comparative effectiveness [e.g., overall survival (OS)] was estimated using several propensity score weighting methods. Resulting real-world parameters informed a cost-effectiveness analysis from the Taiwan payer perspective.
RESULTS: A total of 725 and 240 patients with HCC initiating first-line systemic therapy were classified into the HTA and HTAR cohorts, respectively. Before weighting, the atezo-beva and sorafenib groups differed substantially on markers of tumor burden and liver function. After weighting, the SMDs for these key prognostic variables were balanced. In the HTA cohort, atezo-beva was associated with significantly improved OS but not cost-effectiveness compared with sorafenib (HR: 0.65, 0.48-0.87; ICER: 208,530 USD/QALY). In the HTR cohort, lenvatinib was associated with better OS than atezo-beva in the non-viral hepatitis subgroup.
CONCLUSIONS: LLM-refined hospital EMR data provide uniquely valuable evidence across the HTA/HTAR pipeline, enabling clinically credible target trial emulation and more precise cost-effectiveness estimates. This integrated approach offered a transferable model for embedding real-world EMR evidence into oncology reimbursement decisions.