OBJECTIVES: Despite its potential, the use of Generative AI in health economic modeling is still emerging. Leveraging Generative AI to streamline health economic model development could significantly benefit stakeholders by reducing the time and expertise required. Our objective was to evaluate the feasibility and accuracy of using Generative AI to develop a de novo health economic model.

METHODS: We used hepatitis C treatment as a case study for developing a de novo health economic model. Python was employed to interact with GPT-4 to develop a model structure for the natural history of hepatitis C by searching for published health economic models on PubMed. For parameter estimation, we utilized a multi-agent pipeline and OpenAI libraries. These parameters were then transferred to the R programming language's Heemod package to construct and run the Markov model. All of these steps were packaged into a customized solution, ValueGen.AI. The platform generated health states, transition probabilities, state and treatment costs, discount rates, cycle lengths, and time horizons. We compared the Generative AI-based model's structure, input parameters, and generated outcomes against expert opinion and published studies.

RESULTS: The Generative AI platform constructed a Markov-based model structure for hepatitis C with 10-15 health states and over 22 transitions between states. The platform also provided selected published models used as references. Additionally, the platform estimated values for transition probabilities, costs, and utilities, referencing publications for selected parameters. We repeated the experiment five times and observed variability in the model structure, attributed to variability in the referenced published models. Overall, the generated models had face validity.

CONCLUSIONS: We demonstrate the feasibility of developing a de novo Markov model for a chronic disease using Generative AI. Future research should focus on reducing variability in the model development process and comparing AI-generated model outputs with published models.