Are 237 Published Cost-Effectiveness Models Necessary for Non-Small-Cell Lung Cancer (NSCLC)? Can Open Source Model Platforms Improve Decision-Making and Save Resources?

OBJECTIVES: High costs of collecting evidence for every relevant treatment comparison and patient group over long decision-making time horizons often render economic analysis of trial data impractical. Many healthcare stakeholders use economic models to inform decisions (e.g., drug manufacturers and HTAs). To maximize trust in models, ISPOR promotes model transparency (understandability) and testing face validity, verifying implementation, and assessing predictive accuracy against clinical outcomes.

As part of the Horizon Europe-funded I³Lung Project, we performed an SLR of cost-effectiveness models of NSCLC published between 2012 and 2023, finding 237 unique models. Even accounting for diverse modeling goals, we found this number startling. Leveraging the SLR, we contemplated whether so many individual models are necessary and whether healthcare stakeholders and society could benefit from using customizable open-source models.

METHODS: We assessed heterogeneity in the SLR sample, considering whether these motivated so many unique models, and explored potential benefits that might accrue if open-source models were made available.

RESULTS: Key sources of model heterogeneity were primarily data-related, especially clinical data sources. The structural model elements were homogeneous; 84% used the three-state approach (PFS, PD, death) and most models were Markov. Adherence to best practice guidelines varied, though quality was generally modest and only 20% tested validity.

CONCLUSIONS: For NSCLC, a new paradigm based on open-source models can provide substantial benefits, including standardization around better quality models, better-informed decision-making, enhanced understandability, and reduced duplication. These models should reflect real structural differences needed for different types of analysis (e.g., drug comparisons vs. indirect guided treatment comparisons), necessitating different models. To minimize the models required, features should be comprehensive. Customizability is essential, including pre-programmed distributions and user-friendly entry of regression coefficients. This core model should be verified and validated. While it is unclear who will develop these models, increasing demand (e.g., from journals) for open-source models could drive this.