OBJECTIVES: Generative-AI has demonstrated potential in automating complex tasks through advanced natural language processing. Yet, its application in health economic modeling is still emerging. Partitioned survival models (PSMs) pose unique challenges due to their complexity and need for precise parameter extraction. Leveraging Generative-AI for PSM development could significantly reduce the time and expertise required. This study aimed to evaluate the feasibility and accuracy of using Generative-AI to replicate a PSM, utilizing a published benchmark as a reference.

METHODS: We replicated the PSM from the 2022 Multiple-Myeloma Report by the Institute for Clinical and Economic Review using a life sciences industry-specialized Generative-AI platform, called ValueGen.AI. First, we developed multi-agent pipelines for parameter extraction using GPT-4o and utilizing CrewAI, LangChain, and OpenAI libraries in Python. Second, we created a general PSM template using the Heemod library in R and an API with the Plumber library to automate PSM model creation and execution. We uploaded the published Multiple-Myeloma report to ValueGen.AI and prompted ValueGen.AI to build a PSM for Ide-cel and its comparator (Triple- or Quad Refractor). We evaluated Generative-AI-predicted costs, utilities and Incremental cost-effectiveness ratio (ICER).

RESULTS: The ValueGen.AI platform successfully replicated the model structure and parameters of the published Multiple-Myeloma Report. Generative-AI-based PSM estimated Ide-cel cost at $503,023 (vs $646,000 in the report) and comparator cost at $138,582 (vs $276,000). Generative AI-based PSM estimated Ide-cel QALY at 2.239 (vs 2.24) and comparator QALY at 1.06 (vs 1.08). The error margins for delta cost and delta QALYs between the model generated vs. published values were 1.5% and 1.76%, respectively. Additionally, the estimated ICER value was $308,744 (vs $318,966), resulting in a 3.2% error margin.

CONCLUSIONS: We demonstrate the feasibility of accurately replicating published partitioned survival models using Generative-AI. Future research should aim replicating more complex decision-analytic models, thereby advancing their application in health economic modeling.